1 /* The Blackfin code generation auxiliary output file.
2 Copyright (C) 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
3 Contributed by Analog Devices.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published
9 by the Free Software Foundation; either version 3, or (at your
10 option) any later version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
14 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
15 License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
27 #include "hard-reg-set.h"
29 #include "insn-config.h"
30 #include "insn-codes.h"
31 #include "conditions.h"
32 #include "insn-flags.h"
34 #include "insn-attr.h"
41 #include "target-def.h"
47 #include "integrate.h"
49 #include "langhooks.h"
50 #include "bfin-protos.h"
52 #include "tm-constrs.h"
54 #include "basic-block.h"
55 #include "cfglayout.h"
59 /* A C structure for machine-specific, per-function data.
60 This is added to the cfun structure. */
61 struct machine_function GTY(())
63 int has_hardware_loops;
66 /* Test and compare insns in bfin.md store the information needed to
67 generate branch and scc insns here. */
68 rtx bfin_compare_op0, bfin_compare_op1;
70 /* RTX for condition code flag register and RETS register */
71 extern GTY(()) rtx bfin_cc_rtx;
72 extern GTY(()) rtx bfin_rets_rtx;
73 rtx bfin_cc_rtx, bfin_rets_rtx;
75 int max_arg_registers = 0;
77 /* Arrays used when emitting register names. */
78 const char *short_reg_names[] = SHORT_REGISTER_NAMES;
79 const char *high_reg_names[] = HIGH_REGISTER_NAMES;
80 const char *dregs_pair_names[] = DREGS_PAIR_NAMES;
81 const char *byte_reg_names[] = BYTE_REGISTER_NAMES;
83 static int arg_regs[] = FUNCTION_ARG_REGISTERS;
85 /* Nonzero if -mshared-library-id was given. */
86 static int bfin_lib_id_given;
88 /* Nonzero if -fschedule-insns2 was given. We override it and
89 call the scheduler ourselves during reorg. */
90 static int bfin_flag_schedule_insns2;
92 /* Determines whether we run variable tracking in machine dependent
94 static int bfin_flag_var_tracking;
97 bfin_cpu_t bfin_cpu_type = BFIN_CPU_UNKNOWN;
99 /* -msi-revision support. There are three special values:
100 -1 -msi-revision=none.
101 0xffff -msi-revision=any. */
102 int bfin_si_revision;
104 /* The workarounds enabled */
105 unsigned int bfin_workarounds = 0;
112 unsigned int workarounds;
115 struct bfin_cpu bfin_cpus[] =
117 {"bf522", BFIN_CPU_BF522, 0x0000,
118 WA_SPECULATIVE_LOADS | WA_RETS},
120 {"bf523", BFIN_CPU_BF523, 0x0000,
121 WA_SPECULATIVE_LOADS | WA_RETS},
123 {"bf524", BFIN_CPU_BF524, 0x0000,
124 WA_SPECULATIVE_LOADS | WA_RETS},
126 {"bf525", BFIN_CPU_BF525, 0x0000,
127 WA_SPECULATIVE_LOADS | WA_RETS},
129 {"bf526", BFIN_CPU_BF526, 0x0000,
130 WA_SPECULATIVE_LOADS | WA_RETS},
132 {"bf527", BFIN_CPU_BF527, 0x0000,
133 WA_SPECULATIVE_LOADS | WA_RETS},
135 {"bf531", BFIN_CPU_BF531, 0x0005,
136 WA_SPECULATIVE_LOADS | WA_RETS},
137 {"bf531", BFIN_CPU_BF531, 0x0004,
138 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
139 {"bf531", BFIN_CPU_BF531, 0x0003,
140 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
142 {"bf532", BFIN_CPU_BF532, 0x0005,
143 WA_SPECULATIVE_LOADS | WA_RETS},
144 {"bf532", BFIN_CPU_BF532, 0x0004,
145 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
146 {"bf532", BFIN_CPU_BF532, 0x0003,
147 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
149 {"bf533", BFIN_CPU_BF533, 0x0005,
150 WA_SPECULATIVE_LOADS | WA_RETS},
151 {"bf533", BFIN_CPU_BF533, 0x0004,
152 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
153 {"bf533", BFIN_CPU_BF533, 0x0003,
154 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
156 {"bf534", BFIN_CPU_BF534, 0x0003,
157 WA_SPECULATIVE_LOADS | WA_RETS},
158 {"bf534", BFIN_CPU_BF534, 0x0002,
159 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
160 {"bf534", BFIN_CPU_BF534, 0x0001,
161 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
163 {"bf536", BFIN_CPU_BF536, 0x0003,
164 WA_SPECULATIVE_LOADS | WA_RETS},
165 {"bf536", BFIN_CPU_BF536, 0x0002,
166 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
167 {"bf536", BFIN_CPU_BF536, 0x0001,
168 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
170 {"bf537", BFIN_CPU_BF537, 0x0003,
171 WA_SPECULATIVE_LOADS | WA_RETS},
172 {"bf537", BFIN_CPU_BF537, 0x0002,
173 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
174 {"bf537", BFIN_CPU_BF537, 0x0001,
175 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
177 {"bf538", BFIN_CPU_BF538, 0x0004,
178 WA_SPECULATIVE_LOADS | WA_RETS},
179 {"bf538", BFIN_CPU_BF538, 0x0003,
180 WA_SPECULATIVE_LOADS | WA_RETS},
182 {"bf539", BFIN_CPU_BF539, 0x0004,
183 WA_SPECULATIVE_LOADS | WA_RETS},
184 {"bf539", BFIN_CPU_BF539, 0x0003,
185 WA_SPECULATIVE_LOADS | WA_RETS},
186 {"bf539", BFIN_CPU_BF539, 0x0002,
187 WA_SPECULATIVE_LOADS | WA_RETS},
189 {"bf542", BFIN_CPU_BF542, 0x0000,
190 WA_SPECULATIVE_LOADS | WA_RETS},
192 {"bf544", BFIN_CPU_BF544, 0x0000,
193 WA_SPECULATIVE_LOADS | WA_RETS},
195 {"bf547", BFIN_CPU_BF547, 0x0000,
196 WA_SPECULATIVE_LOADS | WA_RETS},
198 {"bf548", BFIN_CPU_BF548, 0x0000,
199 WA_SPECULATIVE_LOADS | WA_RETS},
201 {"bf549", BFIN_CPU_BF549, 0x0000,
202 WA_SPECULATIVE_LOADS | WA_RETS},
204 {"bf561", BFIN_CPU_BF561, 0x0005, WA_RETS},
205 {"bf561", BFIN_CPU_BF561, 0x0003,
206 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
207 {"bf561", BFIN_CPU_BF561, 0x0002,
208 WA_SPECULATIVE_LOADS | WA_SPECULATIVE_SYNCS | WA_RETS},
213 int splitting_for_sched;
216 bfin_globalize_label (FILE *stream, const char *name)
218 fputs (".global ", stream);
219 assemble_name (stream, name);
225 output_file_start (void)
227 FILE *file = asm_out_file;
230 /* Variable tracking should be run after all optimizations which change order
231 of insns. It also needs a valid CFG. This can't be done in
232 override_options, because flag_var_tracking is finalized after
234 bfin_flag_var_tracking = flag_var_tracking;
235 flag_var_tracking = 0;
237 fprintf (file, ".file \"%s\";\n", input_filename);
239 for (i = 0; arg_regs[i] >= 0; i++)
241 max_arg_registers = i; /* how many arg reg used */
244 /* Called early in the compilation to conditionally modify
245 fixed_regs/call_used_regs. */
248 conditional_register_usage (void)
250 /* initialize condition code flag register rtx */
251 bfin_cc_rtx = gen_rtx_REG (BImode, REG_CC);
252 bfin_rets_rtx = gen_rtx_REG (Pmode, REG_RETS);
255 /* Examine machine-dependent attributes of function type FUNTYPE and return its
256 type. See the definition of E_FUNKIND. */
259 funkind (const_tree funtype)
261 tree attrs = TYPE_ATTRIBUTES (funtype);
262 if (lookup_attribute ("interrupt_handler", attrs))
263 return INTERRUPT_HANDLER;
264 else if (lookup_attribute ("exception_handler", attrs))
265 return EXCPT_HANDLER;
266 else if (lookup_attribute ("nmi_handler", attrs))
272 /* Legitimize PIC addresses. If the address is already position-independent,
273 we return ORIG. Newly generated position-independent addresses go into a
274 reg. This is REG if nonzero, otherwise we allocate register(s) as
275 necessary. PICREG is the register holding the pointer to the PIC offset
279 legitimize_pic_address (rtx orig, rtx reg, rtx picreg)
284 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
289 if (TARGET_ID_SHARED_LIBRARY)
290 unspec = UNSPEC_MOVE_PIC;
291 else if (GET_CODE (addr) == SYMBOL_REF
292 && SYMBOL_REF_FUNCTION_P (addr))
293 unspec = UNSPEC_FUNCDESC_GOT17M4;
295 unspec = UNSPEC_MOVE_FDPIC;
299 gcc_assert (can_create_pseudo_p ());
300 reg = gen_reg_rtx (Pmode);
303 tmp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), unspec);
304 new = gen_const_mem (Pmode, gen_rtx_PLUS (Pmode, picreg, tmp));
306 emit_move_insn (reg, new);
307 if (picreg == pic_offset_table_rtx)
308 crtl->uses_pic_offset_table = 1;
312 else if (GET_CODE (addr) == CONST || GET_CODE (addr) == PLUS)
316 if (GET_CODE (addr) == CONST)
318 addr = XEXP (addr, 0);
319 gcc_assert (GET_CODE (addr) == PLUS);
322 if (XEXP (addr, 0) == picreg)
327 gcc_assert (can_create_pseudo_p ());
328 reg = gen_reg_rtx (Pmode);
331 base = legitimize_pic_address (XEXP (addr, 0), reg, picreg);
332 addr = legitimize_pic_address (XEXP (addr, 1),
333 base == reg ? NULL_RTX : reg,
336 if (GET_CODE (addr) == CONST_INT)
338 gcc_assert (! reload_in_progress && ! reload_completed);
339 addr = force_reg (Pmode, addr);
342 if (GET_CODE (addr) == PLUS && CONSTANT_P (XEXP (addr, 1)))
344 base = gen_rtx_PLUS (Pmode, base, XEXP (addr, 0));
345 addr = XEXP (addr, 1);
348 return gen_rtx_PLUS (Pmode, base, addr);
354 /* Stack frame layout. */
356 /* For a given REGNO, determine whether it must be saved in the function
357 prologue. IS_INTHANDLER specifies whether we're generating a normal
358 prologue or an interrupt/exception one. */
360 must_save_p (bool is_inthandler, unsigned regno)
362 if (D_REGNO_P (regno))
364 bool is_eh_return_reg = false;
365 if (crtl->calls_eh_return)
370 unsigned test = EH_RETURN_DATA_REGNO (j);
371 if (test == INVALID_REGNUM)
374 is_eh_return_reg = true;
378 return (is_eh_return_reg
379 || (df_regs_ever_live_p (regno)
380 && !fixed_regs[regno]
381 && (is_inthandler || !call_used_regs[regno])));
383 else if (P_REGNO_P (regno))
385 return ((df_regs_ever_live_p (regno)
386 && !fixed_regs[regno]
387 && (is_inthandler || !call_used_regs[regno]))
389 && regno == PIC_OFFSET_TABLE_REGNUM
390 && (crtl->uses_pic_offset_table
391 || (TARGET_ID_SHARED_LIBRARY && !current_function_is_leaf))));
394 return ((is_inthandler || !call_used_regs[regno])
395 && (df_regs_ever_live_p (regno)
396 || (!leaf_function_p () && call_used_regs[regno])));
400 /* Compute the number of DREGS to save with a push_multiple operation.
401 This could include registers that aren't modified in the function,
402 since push_multiple only takes a range of registers.
403 If IS_INTHANDLER, then everything that is live must be saved, even
404 if normally call-clobbered.
405 If CONSECUTIVE, return the number of registers we can save in one
406 instruction with a push/pop multiple instruction. */
409 n_dregs_to_save (bool is_inthandler, bool consecutive)
414 for (i = REG_R7 + 1; i-- != REG_R0;)
416 if (must_save_p (is_inthandler, i))
418 else if (consecutive)
424 /* Like n_dregs_to_save, but compute number of PREGS to save. */
427 n_pregs_to_save (bool is_inthandler, bool consecutive)
432 for (i = REG_P5 + 1; i-- != REG_P0;)
433 if (must_save_p (is_inthandler, i))
435 else if (consecutive)
440 /* Determine if we are going to save the frame pointer in the prologue. */
443 must_save_fp_p (void)
445 return frame_pointer_needed || df_regs_ever_live_p (REG_FP);
449 stack_frame_needed_p (void)
451 /* EH return puts a new return address into the frame using an
452 address relative to the frame pointer. */
453 if (crtl->calls_eh_return)
455 return frame_pointer_needed;
458 /* Emit code to save registers in the prologue. SAVEALL is nonzero if we
459 must save all registers; this is used for interrupt handlers.
460 SPREG contains (reg:SI REG_SP). IS_INTHANDLER is true if we're doing
461 this for an interrupt (or exception) handler. */
464 expand_prologue_reg_save (rtx spreg, int saveall, bool is_inthandler)
466 rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
467 rtx predec = gen_rtx_MEM (SImode, predec1);
468 int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
469 int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
470 int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
471 int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
473 int total_consec = ndregs_consec + npregs_consec;
476 if (saveall || is_inthandler)
478 rtx insn = emit_move_insn (predec, gen_rtx_REG (SImode, REG_ASTAT));
479 RTX_FRAME_RELATED_P (insn) = 1;
482 if (total_consec != 0)
485 rtx val = GEN_INT (-total_consec * 4);
486 rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 2));
488 XVECEXP (pat, 0, 0) = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, val),
489 UNSPEC_PUSH_MULTIPLE);
490 XVECEXP (pat, 0, total_consec + 1) = gen_rtx_SET (VOIDmode, spreg,
494 RTX_FRAME_RELATED_P (XVECEXP (pat, 0, total_consec + 1)) = 1;
495 d_to_save = ndregs_consec;
496 dregno = REG_R7 + 1 - ndregs_consec;
497 pregno = REG_P5 + 1 - npregs_consec;
498 for (i = 0; i < total_consec; i++)
500 rtx memref = gen_rtx_MEM (word_mode,
501 gen_rtx_PLUS (Pmode, spreg,
502 GEN_INT (- i * 4 - 4)));
506 subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
512 subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
515 XVECEXP (pat, 0, i + 1) = subpat;
516 RTX_FRAME_RELATED_P (subpat) = 1;
518 insn = emit_insn (pat);
519 RTX_FRAME_RELATED_P (insn) = 1;
522 for (dregno = REG_R0; ndregs != ndregs_consec; dregno++)
524 if (must_save_p (is_inthandler, dregno))
526 rtx insn = emit_move_insn (predec, gen_rtx_REG (word_mode, dregno));
527 RTX_FRAME_RELATED_P (insn) = 1;
531 for (pregno = REG_P0; npregs != npregs_consec; pregno++)
533 if (must_save_p (is_inthandler, pregno))
535 rtx insn = emit_move_insn (predec, gen_rtx_REG (word_mode, pregno));
536 RTX_FRAME_RELATED_P (insn) = 1;
540 for (i = REG_P7 + 1; i < REG_CC; i++)
543 && (df_regs_ever_live_p (i)
544 || (!leaf_function_p () && call_used_regs[i]))))
547 if (i == REG_A0 || i == REG_A1)
548 insn = emit_move_insn (gen_rtx_MEM (PDImode, predec1),
549 gen_rtx_REG (PDImode, i));
551 insn = emit_move_insn (predec, gen_rtx_REG (SImode, i));
552 RTX_FRAME_RELATED_P (insn) = 1;
556 /* Emit code to restore registers in the epilogue. SAVEALL is nonzero if we
557 must save all registers; this is used for interrupt handlers.
558 SPREG contains (reg:SI REG_SP). IS_INTHANDLER is true if we're doing
559 this for an interrupt (or exception) handler. */
562 expand_epilogue_reg_restore (rtx spreg, bool saveall, bool is_inthandler)
564 rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
565 rtx postinc = gen_rtx_MEM (SImode, postinc1);
567 int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
568 int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
569 int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
570 int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
571 int total_consec = ndregs_consec + npregs_consec;
575 /* A slightly crude technique to stop flow from trying to delete "dead"
577 MEM_VOLATILE_P (postinc) = 1;
579 for (i = REG_CC - 1; i > REG_P7; i--)
582 && (df_regs_ever_live_p (i)
583 || (!leaf_function_p () && call_used_regs[i]))))
585 if (i == REG_A0 || i == REG_A1)
587 rtx mem = gen_rtx_MEM (PDImode, postinc1);
588 MEM_VOLATILE_P (mem) = 1;
589 emit_move_insn (gen_rtx_REG (PDImode, i), mem);
592 emit_move_insn (gen_rtx_REG (SImode, i), postinc);
595 regno = REG_P5 - npregs_consec;
596 for (; npregs != npregs_consec; regno--)
598 if (must_save_p (is_inthandler, regno))
600 emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
604 regno = REG_R7 - ndregs_consec;
605 for (; ndregs != ndregs_consec; regno--)
607 if (must_save_p (is_inthandler, regno))
609 emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
614 if (total_consec != 0)
616 rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 1));
618 = gen_rtx_SET (VOIDmode, spreg,
619 gen_rtx_PLUS (Pmode, spreg,
620 GEN_INT (total_consec * 4)));
622 if (npregs_consec > 0)
627 for (i = 0; i < total_consec; i++)
630 ? gen_rtx_PLUS (Pmode, spreg, GEN_INT (i * 4))
632 rtx memref = gen_rtx_MEM (word_mode, addr);
635 XVECEXP (pat, 0, i + 1)
636 = gen_rtx_SET (VOIDmode, gen_rtx_REG (word_mode, regno), memref);
638 if (npregs_consec > 0)
640 if (--npregs_consec == 0)
645 insn = emit_insn (pat);
646 RTX_FRAME_RELATED_P (insn) = 1;
648 if (saveall || is_inthandler)
649 emit_move_insn (gen_rtx_REG (SImode, REG_ASTAT), postinc);
652 /* Perform any needed actions needed for a function that is receiving a
653 variable number of arguments.
657 MODE and TYPE are the mode and type of the current parameter.
659 PRETEND_SIZE is a variable that should be set to the amount of stack
660 that must be pushed by the prolog to pretend that our caller pushed
663 Normally, this macro will push all remaining incoming registers on the
664 stack and set PRETEND_SIZE to the length of the registers pushed.
667 - VDSP C compiler manual (our ABI) says that a variable args function
668 should save the R0, R1 and R2 registers in the stack.
669 - The caller will always leave space on the stack for the
670 arguments that are passed in registers, so we dont have
671 to leave any extra space.
672 - now, the vastart pointer can access all arguments from the stack. */
675 setup_incoming_varargs (CUMULATIVE_ARGS *cum,
676 enum machine_mode mode ATTRIBUTE_UNUSED,
677 tree type ATTRIBUTE_UNUSED, int *pretend_size,
686 /* The move for named arguments will be generated automatically by the
687 compiler. We need to generate the move rtx for the unnamed arguments
688 if they are in the first 3 words. We assume at least 1 named argument
689 exists, so we never generate [ARGP] = R0 here. */
691 for (i = cum->words + 1; i < max_arg_registers; i++)
693 mem = gen_rtx_MEM (Pmode,
694 plus_constant (arg_pointer_rtx, (i * UNITS_PER_WORD)));
695 emit_move_insn (mem, gen_rtx_REG (Pmode, i));
701 /* Value should be nonzero if functions must have frame pointers.
702 Zero means the frame pointer need not be set up (and parms may
703 be accessed via the stack pointer) in functions that seem suitable. */
706 bfin_frame_pointer_required (void)
708 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
710 if (fkind != SUBROUTINE)
713 /* We turn on -fomit-frame-pointer if -momit-leaf-frame-pointer is used,
714 so we have to override it for non-leaf functions. */
715 if (TARGET_OMIT_LEAF_FRAME_POINTER && ! current_function_is_leaf)
721 /* Return the number of registers pushed during the prologue. */
724 n_regs_saved_by_prologue (void)
726 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
727 bool is_inthandler = fkind != SUBROUTINE;
728 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
729 bool all = (lookup_attribute ("saveall", attrs) != NULL_TREE
730 || (is_inthandler && !current_function_is_leaf));
731 int ndregs = all ? 8 : n_dregs_to_save (is_inthandler, false);
732 int npregs = all ? 6 : n_pregs_to_save (is_inthandler, false);
733 int n = ndregs + npregs;
736 if (all || stack_frame_needed_p ())
737 /* We use a LINK instruction in this case. */
741 if (must_save_fp_p ())
743 if (! current_function_is_leaf)
747 if (fkind != SUBROUTINE || all)
748 /* Increment once for ASTAT. */
751 if (fkind != SUBROUTINE)
754 if (lookup_attribute ("nesting", attrs))
758 for (i = REG_P7 + 1; i < REG_CC; i++)
760 || (fkind != SUBROUTINE
761 && (df_regs_ever_live_p (i)
762 || (!leaf_function_p () && call_used_regs[i]))))
763 n += i == REG_A0 || i == REG_A1 ? 2 : 1;
768 /* Return the offset between two registers, one to be eliminated, and the other
769 its replacement, at the start of a routine. */
772 bfin_initial_elimination_offset (int from, int to)
774 HOST_WIDE_INT offset = 0;
776 if (from == ARG_POINTER_REGNUM)
777 offset = n_regs_saved_by_prologue () * 4;
779 if (to == STACK_POINTER_REGNUM)
781 if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
782 offset += crtl->outgoing_args_size;
783 else if (crtl->outgoing_args_size)
784 offset += FIXED_STACK_AREA;
786 offset += get_frame_size ();
792 /* Emit code to load a constant CONSTANT into register REG; setting
793 RTX_FRAME_RELATED_P on all insns we generate if RELATED is true.
794 Make sure that the insns we generate need not be split. */
797 frame_related_constant_load (rtx reg, HOST_WIDE_INT constant, bool related)
800 rtx cst = GEN_INT (constant);
802 if (constant >= -32768 && constant < 65536)
803 insn = emit_move_insn (reg, cst);
806 /* We don't call split_load_immediate here, since dwarf2out.c can get
807 confused about some of the more clever sequences it can generate. */
808 insn = emit_insn (gen_movsi_high (reg, cst));
810 RTX_FRAME_RELATED_P (insn) = 1;
811 insn = emit_insn (gen_movsi_low (reg, reg, cst));
814 RTX_FRAME_RELATED_P (insn) = 1;
817 /* Generate efficient code to add a value to a P register.
818 Set RTX_FRAME_RELATED_P on the generated insns if FRAME is nonzero.
819 EPILOGUE_P is zero if this function is called for prologue,
820 otherwise it's nonzero. And it's less than zero if this is for
824 add_to_reg (rtx reg, HOST_WIDE_INT value, int frame, int epilogue_p)
829 /* Choose whether to use a sequence using a temporary register, or
830 a sequence with multiple adds. We can add a signed 7-bit value
831 in one instruction. */
832 if (value > 120 || value < -120)
840 /* For prologue or normal epilogue, P1 can be safely used
841 as the temporary register. For sibcall epilogue, we try to find
842 a call used P register, which will be restored in epilogue.
843 If we cannot find such a P register, we have to use one I register
847 tmpreg = gen_rtx_REG (SImode, REG_P1);
851 for (i = REG_P0; i <= REG_P5; i++)
852 if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
854 && i == PIC_OFFSET_TABLE_REGNUM
855 && (crtl->uses_pic_offset_table
856 || (TARGET_ID_SHARED_LIBRARY
857 && ! current_function_is_leaf))))
860 tmpreg = gen_rtx_REG (SImode, i);
863 tmpreg = gen_rtx_REG (SImode, REG_P1);
864 tmpreg2 = gen_rtx_REG (SImode, REG_I0);
865 emit_move_insn (tmpreg2, tmpreg);
870 frame_related_constant_load (tmpreg, value, TRUE);
872 insn = emit_move_insn (tmpreg, GEN_INT (value));
874 insn = emit_insn (gen_addsi3 (reg, reg, tmpreg));
876 RTX_FRAME_RELATED_P (insn) = 1;
878 if (tmpreg2 != NULL_RTX)
879 emit_move_insn (tmpreg, tmpreg2);
890 /* We could use -62, but that would leave the stack unaligned, so
894 insn = emit_insn (gen_addsi3 (reg, reg, GEN_INT (size)));
896 RTX_FRAME_RELATED_P (insn) = 1;
902 /* Generate a LINK insn for a frame sized FRAME_SIZE. If this constant
903 is too large, generate a sequence of insns that has the same effect.
904 SPREG contains (reg:SI REG_SP). */
907 emit_link_insn (rtx spreg, HOST_WIDE_INT frame_size)
909 HOST_WIDE_INT link_size = frame_size;
913 if (link_size > 262140)
916 /* Use a LINK insn with as big a constant as possible, then subtract
917 any remaining size from the SP. */
918 insn = emit_insn (gen_link (GEN_INT (-8 - link_size)));
919 RTX_FRAME_RELATED_P (insn) = 1;
921 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
923 rtx set = XVECEXP (PATTERN (insn), 0, i);
924 gcc_assert (GET_CODE (set) == SET);
925 RTX_FRAME_RELATED_P (set) = 1;
928 frame_size -= link_size;
932 /* Must use a call-clobbered PREG that isn't the static chain. */
933 rtx tmpreg = gen_rtx_REG (Pmode, REG_P1);
935 frame_related_constant_load (tmpreg, -frame_size, TRUE);
936 insn = emit_insn (gen_addsi3 (spreg, spreg, tmpreg));
937 RTX_FRAME_RELATED_P (insn) = 1;
941 /* Return the number of bytes we must reserve for outgoing arguments
942 in the current function's stack frame. */
947 if (crtl->outgoing_args_size)
949 if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
950 return crtl->outgoing_args_size;
952 return FIXED_STACK_AREA;
957 /* Save RETS and FP, and allocate a stack frame. ALL is true if the
958 function must save all its registers (true only for certain interrupt
962 do_link (rtx spreg, HOST_WIDE_INT frame_size, bool all)
964 frame_size += arg_area_size ();
966 if (all || stack_frame_needed_p ()
967 || (must_save_fp_p () && ! current_function_is_leaf))
968 emit_link_insn (spreg, frame_size);
971 if (! current_function_is_leaf)
973 rtx pat = gen_movsi (gen_rtx_MEM (Pmode,
974 gen_rtx_PRE_DEC (Pmode, spreg)),
976 rtx insn = emit_insn (pat);
977 RTX_FRAME_RELATED_P (insn) = 1;
979 if (must_save_fp_p ())
981 rtx pat = gen_movsi (gen_rtx_MEM (Pmode,
982 gen_rtx_PRE_DEC (Pmode, spreg)),
983 gen_rtx_REG (Pmode, REG_FP));
984 rtx insn = emit_insn (pat);
985 RTX_FRAME_RELATED_P (insn) = 1;
987 add_to_reg (spreg, -frame_size, 1, 0);
991 /* Like do_link, but used for epilogues to deallocate the stack frame.
992 EPILOGUE_P is zero if this function is called for prologue,
993 otherwise it's nonzero. And it's less than zero if this is for
997 do_unlink (rtx spreg, HOST_WIDE_INT frame_size, bool all, int epilogue_p)
999 frame_size += arg_area_size ();
1001 if (all || stack_frame_needed_p ())
1002 emit_insn (gen_unlink ());
1005 rtx postinc = gen_rtx_MEM (Pmode, gen_rtx_POST_INC (Pmode, spreg));
1007 add_to_reg (spreg, frame_size, 0, epilogue_p);
1008 if (must_save_fp_p ())
1010 rtx fpreg = gen_rtx_REG (Pmode, REG_FP);
1011 emit_move_insn (fpreg, postinc);
1012 emit_insn (gen_rtx_USE (VOIDmode, fpreg));
1014 if (! current_function_is_leaf)
1016 emit_move_insn (bfin_rets_rtx, postinc);
1017 emit_insn (gen_rtx_USE (VOIDmode, bfin_rets_rtx));
1022 /* Generate a prologue suitable for a function of kind FKIND. This is
1023 called for interrupt and exception handler prologues.
1024 SPREG contains (reg:SI REG_SP). */
1027 expand_interrupt_handler_prologue (rtx spreg, e_funkind fkind, bool all)
1029 HOST_WIDE_INT frame_size = get_frame_size ();
1030 rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
1031 rtx predec = gen_rtx_MEM (SImode, predec1);
1033 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1034 tree kspisusp = lookup_attribute ("kspisusp", attrs);
1038 insn = emit_move_insn (spreg, gen_rtx_REG (Pmode, REG_USP));
1039 RTX_FRAME_RELATED_P (insn) = 1;
1042 /* We need space on the stack in case we need to save the argument
1044 if (fkind == EXCPT_HANDLER)
1046 insn = emit_insn (gen_addsi3 (spreg, spreg, GEN_INT (-12)));
1047 RTX_FRAME_RELATED_P (insn) = 1;
1050 /* If we're calling other functions, they won't save their call-clobbered
1051 registers, so we must save everything here. */
1052 if (!current_function_is_leaf)
1054 expand_prologue_reg_save (spreg, all, true);
1056 if (lookup_attribute ("nesting", attrs))
1058 rtx srcreg = gen_rtx_REG (Pmode, (fkind == EXCPT_HANDLER ? REG_RETX
1059 : fkind == NMI_HANDLER ? REG_RETN
1061 insn = emit_move_insn (predec, srcreg);
1062 RTX_FRAME_RELATED_P (insn) = 1;
1065 do_link (spreg, frame_size, all);
1067 if (fkind == EXCPT_HANDLER)
1069 rtx r0reg = gen_rtx_REG (SImode, REG_R0);
1070 rtx r1reg = gen_rtx_REG (SImode, REG_R1);
1071 rtx r2reg = gen_rtx_REG (SImode, REG_R2);
1074 insn = emit_move_insn (r0reg, gen_rtx_REG (SImode, REG_SEQSTAT));
1075 insn = emit_insn (gen_ashrsi3 (r0reg, r0reg, GEN_INT (26)));
1076 insn = emit_insn (gen_ashlsi3 (r0reg, r0reg, GEN_INT (26)));
1077 insn = emit_move_insn (r1reg, spreg);
1078 insn = emit_move_insn (r2reg, gen_rtx_REG (Pmode, REG_FP));
1079 insn = emit_insn (gen_addsi3 (r2reg, r2reg, GEN_INT (8)));
1083 /* Generate an epilogue suitable for a function of kind FKIND. This is
1084 called for interrupt and exception handler epilogues.
1085 SPREG contains (reg:SI REG_SP). */
1088 expand_interrupt_handler_epilogue (rtx spreg, e_funkind fkind, bool all)
1090 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1091 rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
1092 rtx postinc = gen_rtx_MEM (SImode, postinc1);
1094 /* A slightly crude technique to stop flow from trying to delete "dead"
1096 MEM_VOLATILE_P (postinc) = 1;
1098 do_unlink (spreg, get_frame_size (), all, 1);
1100 if (lookup_attribute ("nesting", attrs))
1102 rtx srcreg = gen_rtx_REG (Pmode, (fkind == EXCPT_HANDLER ? REG_RETX
1103 : fkind == NMI_HANDLER ? REG_RETN
1105 emit_move_insn (srcreg, postinc);
1108 /* If we're calling other functions, they won't save their call-clobbered
1109 registers, so we must save (and restore) everything here. */
1110 if (!current_function_is_leaf)
1113 expand_epilogue_reg_restore (spreg, all, true);
1115 /* Deallocate any space we left on the stack in case we needed to save the
1116 argument registers. */
1117 if (fkind == EXCPT_HANDLER)
1118 emit_insn (gen_addsi3 (spreg, spreg, GEN_INT (12)));
1120 emit_jump_insn (gen_return_internal (GEN_INT (fkind)));
1123 /* Used while emitting the prologue to generate code to load the correct value
1124 into the PIC register, which is passed in DEST. */
1127 bfin_load_pic_reg (rtx dest)
1129 struct cgraph_local_info *i = NULL;
1132 if (flag_unit_at_a_time)
1133 i = cgraph_local_info (current_function_decl);
1135 /* Functions local to the translation unit don't need to reload the
1136 pic reg, since the caller always passes a usable one. */
1138 return pic_offset_table_rtx;
1140 if (bfin_lib_id_given)
1141 addr = plus_constant (pic_offset_table_rtx, -4 - bfin_library_id * 4);
1143 addr = gen_rtx_PLUS (Pmode, pic_offset_table_rtx,
1144 gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
1145 UNSPEC_LIBRARY_OFFSET));
1146 insn = emit_insn (gen_movsi (dest, gen_rtx_MEM (Pmode, addr)));
1150 /* Generate RTL for the prologue of the current function. */
1153 bfin_expand_prologue (void)
1155 HOST_WIDE_INT frame_size = get_frame_size ();
1156 rtx spreg = gen_rtx_REG (Pmode, REG_SP);
1157 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
1158 rtx pic_reg_loaded = NULL_RTX;
1159 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1160 bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
1162 if (fkind != SUBROUTINE)
1164 expand_interrupt_handler_prologue (spreg, fkind, all);
1168 if (crtl->limit_stack
1169 || TARGET_STACK_CHECK_L1)
1171 HOST_WIDE_INT offset
1172 = bfin_initial_elimination_offset (ARG_POINTER_REGNUM,
1173 STACK_POINTER_REGNUM);
1174 rtx lim = crtl->limit_stack ? stack_limit_rtx : NULL_RTX;
1175 rtx p2reg = gen_rtx_REG (Pmode, REG_P2);
1179 emit_move_insn (p2reg, gen_int_mode (0xFFB00000, SImode));
1180 emit_move_insn (p2reg, gen_rtx_MEM (Pmode, p2reg));
1183 if (GET_CODE (lim) == SYMBOL_REF)
1185 if (TARGET_ID_SHARED_LIBRARY)
1187 rtx p1reg = gen_rtx_REG (Pmode, REG_P1);
1189 pic_reg_loaded = bfin_load_pic_reg (p2reg);
1190 val = legitimize_pic_address (stack_limit_rtx, p1reg,
1192 emit_move_insn (p1reg, val);
1193 frame_related_constant_load (p2reg, offset, FALSE);
1194 emit_insn (gen_addsi3 (p2reg, p2reg, p1reg));
1199 rtx limit = plus_constant (lim, offset);
1200 emit_move_insn (p2reg, limit);
1207 emit_move_insn (p2reg, lim);
1208 add_to_reg (p2reg, offset, 0, 0);
1211 emit_insn (gen_compare_lt (bfin_cc_rtx, spreg, lim));
1212 emit_insn (gen_trapifcc ());
1214 expand_prologue_reg_save (spreg, all, false);
1216 do_link (spreg, frame_size, false);
1218 if (TARGET_ID_SHARED_LIBRARY
1220 && (crtl->uses_pic_offset_table
1221 || !current_function_is_leaf))
1222 bfin_load_pic_reg (pic_offset_table_rtx);
1225 /* Generate RTL for the epilogue of the current function. NEED_RETURN is zero
1226 if this is for a sibcall. EH_RETURN is nonzero if we're expanding an
1227 eh_return pattern. SIBCALL_P is true if this is a sibcall epilogue,
1231 bfin_expand_epilogue (int need_return, int eh_return, bool sibcall_p)
1233 rtx spreg = gen_rtx_REG (Pmode, REG_SP);
1234 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
1235 int e = sibcall_p ? -1 : 1;
1236 tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
1237 bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
1239 if (fkind != SUBROUTINE)
1241 expand_interrupt_handler_epilogue (spreg, fkind, all);
1245 do_unlink (spreg, get_frame_size (), false, e);
1247 expand_epilogue_reg_restore (spreg, all, false);
1249 /* Omit the return insn if this is for a sibcall. */
1254 emit_insn (gen_addsi3 (spreg, spreg, gen_rtx_REG (Pmode, REG_P2)));
1256 emit_jump_insn (gen_return_internal (GEN_INT (SUBROUTINE)));
1259 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
1262 bfin_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED,
1263 unsigned int new_reg)
1265 /* Interrupt functions can only use registers that have already been
1266 saved by the prologue, even if they would normally be
1269 if (funkind (TREE_TYPE (current_function_decl)) != SUBROUTINE
1270 && !df_regs_ever_live_p (new_reg))
1276 /* Return the value of the return address for the frame COUNT steps up
1277 from the current frame, after the prologue.
1278 We punt for everything but the current frame by returning const0_rtx. */
1281 bfin_return_addr_rtx (int count)
1286 return get_hard_reg_initial_val (Pmode, REG_RETS);
1289 /* Try machine-dependent ways of modifying an illegitimate address X
1290 to be legitimate. If we find one, return the new, valid address,
1291 otherwise return NULL_RTX.
1293 OLDX is the address as it was before break_out_memory_refs was called.
1294 In some cases it is useful to look at this to decide what needs to be done.
1296 MODE is the mode of the memory reference. */
1299 legitimize_address (rtx x ATTRIBUTE_UNUSED, rtx oldx ATTRIBUTE_UNUSED,
1300 enum machine_mode mode ATTRIBUTE_UNUSED)
1306 bfin_delegitimize_address (rtx orig_x)
1310 if (GET_CODE (x) != MEM)
1314 if (GET_CODE (x) == PLUS
1315 && GET_CODE (XEXP (x, 1)) == UNSPEC
1316 && XINT (XEXP (x, 1), 1) == UNSPEC_MOVE_PIC
1317 && GET_CODE (XEXP (x, 0)) == REG
1318 && REGNO (XEXP (x, 0)) == PIC_OFFSET_TABLE_REGNUM)
1319 return XVECEXP (XEXP (x, 1), 0, 0);
1324 /* This predicate is used to compute the length of a load/store insn.
1325 OP is a MEM rtx, we return nonzero if its addressing mode requires a
1326 32-bit instruction. */
1329 effective_address_32bit_p (rtx op, enum machine_mode mode)
1331 HOST_WIDE_INT offset;
1333 mode = GET_MODE (op);
1336 if (GET_CODE (op) != PLUS)
1338 gcc_assert (REG_P (op) || GET_CODE (op) == POST_INC
1339 || GET_CODE (op) == PRE_DEC || GET_CODE (op) == POST_DEC);
1343 if (GET_CODE (XEXP (op, 1)) == UNSPEC)
1346 offset = INTVAL (XEXP (op, 1));
1348 /* All byte loads use a 16-bit offset. */
1349 if (GET_MODE_SIZE (mode) == 1)
1352 if (GET_MODE_SIZE (mode) == 4)
1354 /* Frame pointer relative loads can use a negative offset, all others
1355 are restricted to a small positive one. */
1356 if (XEXP (op, 0) == frame_pointer_rtx)
1357 return offset < -128 || offset > 60;
1358 return offset < 0 || offset > 60;
1361 /* Must be HImode now. */
1362 return offset < 0 || offset > 30;
1365 /* Returns true if X is a memory reference using an I register. */
1367 bfin_dsp_memref_p (rtx x)
1372 if (GET_CODE (x) == POST_INC || GET_CODE (x) == PRE_INC
1373 || GET_CODE (x) == POST_DEC || GET_CODE (x) == PRE_DEC)
1378 /* Return cost of the memory address ADDR.
1379 All addressing modes are equally cheap on the Blackfin. */
1382 bfin_address_cost (rtx addr ATTRIBUTE_UNUSED)
1387 /* Subroutine of print_operand; used to print a memory reference X to FILE. */
1390 print_address_operand (FILE *file, rtx x)
1392 switch (GET_CODE (x))
1395 output_address (XEXP (x, 0));
1396 fprintf (file, "+");
1397 output_address (XEXP (x, 1));
1401 fprintf (file, "--");
1402 output_address (XEXP (x, 0));
1405 output_address (XEXP (x, 0));
1406 fprintf (file, "++");
1409 output_address (XEXP (x, 0));
1410 fprintf (file, "--");
1414 gcc_assert (GET_CODE (x) != MEM);
1415 print_operand (file, x, 0);
1420 /* Adding intp DImode support by Tony
1426 print_operand (FILE *file, rtx x, char code)
1428 enum machine_mode mode;
1432 if (GET_MODE (current_output_insn) == SImode)
1433 fprintf (file, " ||");
1435 fprintf (file, ";");
1439 mode = GET_MODE (x);
1444 switch (GET_CODE (x))
1447 fprintf (file, "e");
1450 fprintf (file, "ne");
1453 fprintf (file, "g");
1456 fprintf (file, "l");
1459 fprintf (file, "ge");
1462 fprintf (file, "le");
1465 fprintf (file, "g");
1468 fprintf (file, "l");
1471 fprintf (file, "ge");
1474 fprintf (file, "le");
1477 output_operand_lossage ("invalid %%j value");
1481 case 'J': /* reverse logic */
1482 switch (GET_CODE(x))
1485 fprintf (file, "ne");
1488 fprintf (file, "e");
1491 fprintf (file, "le");
1494 fprintf (file, "ge");
1497 fprintf (file, "l");
1500 fprintf (file, "g");
1503 fprintf (file, "le");
1506 fprintf (file, "ge");
1509 fprintf (file, "l");
1512 fprintf (file, "g");
1515 output_operand_lossage ("invalid %%J value");
1520 switch (GET_CODE (x))
1526 fprintf (file, "%s", short_reg_names[REGNO (x)]);
1528 output_operand_lossage ("invalid operand for code '%c'", code);
1530 else if (code == 'd')
1533 fprintf (file, "%s", high_reg_names[REGNO (x)]);
1535 output_operand_lossage ("invalid operand for code '%c'", code);
1537 else if (code == 'w')
1539 if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
1540 fprintf (file, "%s.w", reg_names[REGNO (x)]);
1542 output_operand_lossage ("invalid operand for code '%c'", code);
1544 else if (code == 'x')
1546 if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
1547 fprintf (file, "%s.x", reg_names[REGNO (x)]);
1549 output_operand_lossage ("invalid operand for code '%c'", code);
1551 else if (code == 'v')
1553 if (REGNO (x) == REG_A0)
1554 fprintf (file, "AV0");
1555 else if (REGNO (x) == REG_A1)
1556 fprintf (file, "AV1");
1558 output_operand_lossage ("invalid operand for code '%c'", code);
1560 else if (code == 'D')
1562 if (D_REGNO_P (REGNO (x)))
1563 fprintf (file, "%s", dregs_pair_names[REGNO (x)]);
1565 output_operand_lossage ("invalid operand for code '%c'", code);
1567 else if (code == 'H')
1569 if ((mode == DImode || mode == DFmode) && REG_P (x))
1570 fprintf (file, "%s", reg_names[REGNO (x) + 1]);
1572 output_operand_lossage ("invalid operand for code '%c'", code);
1574 else if (code == 'T')
1576 if (D_REGNO_P (REGNO (x)))
1577 fprintf (file, "%s", byte_reg_names[REGNO (x)]);
1579 output_operand_lossage ("invalid operand for code '%c'", code);
1582 fprintf (file, "%s", reg_names[REGNO (x)]);
1588 print_address_operand (file, x);
1600 fputs ("(FU)", file);
1603 fputs ("(T)", file);
1606 fputs ("(TFU)", file);
1609 fputs ("(W32)", file);
1612 fputs ("(IS)", file);
1615 fputs ("(IU)", file);
1618 fputs ("(IH)", file);
1621 fputs ("(M)", file);
1624 fputs ("(IS,M)", file);
1627 fputs ("(ISS2)", file);
1630 fputs ("(S2RND)", file);
1637 else if (code == 'b')
1639 if (INTVAL (x) == 0)
1641 else if (INTVAL (x) == 1)
1647 /* Moves to half registers with d or h modifiers always use unsigned
1649 else if (code == 'd')
1650 x = GEN_INT ((INTVAL (x) >> 16) & 0xffff);
1651 else if (code == 'h')
1652 x = GEN_INT (INTVAL (x) & 0xffff);
1653 else if (code == 'N')
1654 x = GEN_INT (-INTVAL (x));
1655 else if (code == 'X')
1656 x = GEN_INT (exact_log2 (0xffffffff & INTVAL (x)));
1657 else if (code == 'Y')
1658 x = GEN_INT (exact_log2 (0xffffffff & ~INTVAL (x)));
1659 else if (code == 'Z')
1660 /* Used for LINK insns. */
1661 x = GEN_INT (-8 - INTVAL (x));
1666 output_addr_const (file, x);
1670 output_operand_lossage ("invalid const_double operand");
1674 switch (XINT (x, 1))
1676 case UNSPEC_MOVE_PIC:
1677 output_addr_const (file, XVECEXP (x, 0, 0));
1678 fprintf (file, "@GOT");
1681 case UNSPEC_MOVE_FDPIC:
1682 output_addr_const (file, XVECEXP (x, 0, 0));
1683 fprintf (file, "@GOT17M4");
1686 case UNSPEC_FUNCDESC_GOT17M4:
1687 output_addr_const (file, XVECEXP (x, 0, 0));
1688 fprintf (file, "@FUNCDESC_GOT17M4");
1691 case UNSPEC_LIBRARY_OFFSET:
1692 fprintf (file, "_current_shared_library_p5_offset_");
1701 output_addr_const (file, x);
1706 /* Argument support functions. */
1708 /* Initialize a variable CUM of type CUMULATIVE_ARGS
1709 for a call to a function whose data type is FNTYPE.
1710 For a library call, FNTYPE is 0.
1711 VDSP C Compiler manual, our ABI says that
1712 first 3 words of arguments will use R0, R1 and R2.
1716 init_cumulative_args (CUMULATIVE_ARGS *cum, tree fntype,
1717 rtx libname ATTRIBUTE_UNUSED)
1719 static CUMULATIVE_ARGS zero_cum;
1723 /* Set up the number of registers to use for passing arguments. */
1725 cum->nregs = max_arg_registers;
1726 cum->arg_regs = arg_regs;
1728 cum->call_cookie = CALL_NORMAL;
1729 /* Check for a longcall attribute. */
1730 if (fntype && lookup_attribute ("shortcall", TYPE_ATTRIBUTES (fntype)))
1731 cum->call_cookie |= CALL_SHORT;
1732 else if (fntype && lookup_attribute ("longcall", TYPE_ATTRIBUTES (fntype)))
1733 cum->call_cookie |= CALL_LONG;
1738 /* Update the data in CUM to advance over an argument
1739 of mode MODE and data type TYPE.
1740 (TYPE is null for libcalls where that information may not be available.) */
1743 function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type,
1744 int named ATTRIBUTE_UNUSED)
1746 int count, bytes, words;
1748 bytes = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
1749 words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
1751 cum->words += words;
1752 cum->nregs -= words;
1754 if (cum->nregs <= 0)
1757 cum->arg_regs = NULL;
1761 for (count = 1; count <= words; count++)
1768 /* Define where to put the arguments to a function.
1769 Value is zero to push the argument on the stack,
1770 or a hard register in which to store the argument.
1772 MODE is the argument's machine mode.
1773 TYPE is the data type of the argument (as a tree).
1774 This is null for libcalls where that information may
1776 CUM is a variable of type CUMULATIVE_ARGS which gives info about
1777 the preceding args and about the function being called.
1778 NAMED is nonzero if this argument is a named parameter
1779 (otherwise it is an extra parameter matching an ellipsis). */
1782 function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type,
1783 int named ATTRIBUTE_UNUSED)
1786 = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
1788 if (mode == VOIDmode)
1789 /* Compute operand 2 of the call insn. */
1790 return GEN_INT (cum->call_cookie);
1796 return gen_rtx_REG (mode, *(cum->arg_regs));
1801 /* For an arg passed partly in registers and partly in memory,
1802 this is the number of bytes passed in registers.
1803 For args passed entirely in registers or entirely in memory, zero.
1805 Refer VDSP C Compiler manual, our ABI.
1806 First 3 words are in registers. So, if an argument is larger
1807 than the registers available, it will span the register and
1811 bfin_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
1812 tree type ATTRIBUTE_UNUSED,
1813 bool named ATTRIBUTE_UNUSED)
1816 = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
1817 int bytes_left = cum->nregs * UNITS_PER_WORD;
1822 if (bytes_left == 0)
1824 if (bytes > bytes_left)
1829 /* Variable sized types are passed by reference. */
1832 bfin_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
1833 enum machine_mode mode ATTRIBUTE_UNUSED,
1834 const_tree type, bool named ATTRIBUTE_UNUSED)
1836 return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST;
1839 /* Decide whether a type should be returned in memory (true)
1840 or in a register (false). This is called by the macro
1841 TARGET_RETURN_IN_MEMORY. */
1844 bfin_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
1846 int size = int_size_in_bytes (type);
1847 return size > 2 * UNITS_PER_WORD || size == -1;
1850 /* Register in which address to store a structure value
1851 is passed to a function. */
1853 bfin_struct_value_rtx (tree fntype ATTRIBUTE_UNUSED,
1854 int incoming ATTRIBUTE_UNUSED)
1856 return gen_rtx_REG (Pmode, REG_P0);
1859 /* Return true when register may be used to pass function parameters. */
1862 function_arg_regno_p (int n)
1865 for (i = 0; arg_regs[i] != -1; i++)
1866 if (n == arg_regs[i])
1871 /* Returns 1 if OP contains a symbol reference */
1874 symbolic_reference_mentioned_p (rtx op)
1876 register const char *fmt;
1879 if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF)
1882 fmt = GET_RTX_FORMAT (GET_CODE (op));
1883 for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--)
1889 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
1890 if (symbolic_reference_mentioned_p (XVECEXP (op, i, j)))
1894 else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i)))
1901 /* Decide whether we can make a sibling call to a function. DECL is the
1902 declaration of the function being targeted by the call and EXP is the
1903 CALL_EXPR representing the call. */
1906 bfin_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED,
1907 tree exp ATTRIBUTE_UNUSED)
1909 e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
1910 if (fkind != SUBROUTINE)
1912 if (!TARGET_ID_SHARED_LIBRARY || TARGET_SEP_DATA)
1915 /* When compiling for ID shared libraries, can't sibcall a local function
1916 from a non-local function, because the local function thinks it does
1917 not need to reload P5 in the prologue, but the sibcall wil pop P5 in the
1918 sibcall epilogue, and we end up with the wrong value in P5. */
1920 if (!flag_unit_at_a_time || decl == NULL)
1921 /* Not enough information. */
1925 struct cgraph_local_info *this_func, *called_func;
1927 this_func = cgraph_local_info (current_function_decl);
1928 called_func = cgraph_local_info (decl);
1929 return !called_func->local || this_func->local;
1933 /* Emit RTL insns to initialize the variable parts of a trampoline at
1934 TRAMP. FNADDR is an RTX for the address of the function's pure
1935 code. CXT is an RTX for the static chain value for the function. */
1938 initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
1940 rtx t1 = copy_to_reg (fnaddr);
1941 rtx t2 = copy_to_reg (cxt);
1947 rtx a = memory_address (Pmode, plus_constant (tramp, 8));
1948 addr = memory_address (Pmode, tramp);
1949 emit_move_insn (gen_rtx_MEM (SImode, addr), a);
1953 addr = memory_address (Pmode, plus_constant (tramp, i + 2));
1954 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t1));
1955 emit_insn (gen_ashrsi3 (t1, t1, GEN_INT (16)));
1956 addr = memory_address (Pmode, plus_constant (tramp, i + 6));
1957 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t1));
1959 addr = memory_address (Pmode, plus_constant (tramp, i + 10));
1960 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t2));
1961 emit_insn (gen_ashrsi3 (t2, t2, GEN_INT (16)));
1962 addr = memory_address (Pmode, plus_constant (tramp, i + 14));
1963 emit_move_insn (gen_rtx_MEM (HImode, addr), gen_lowpart (HImode, t2));
1966 /* Emit insns to move operands[1] into operands[0]. */
1969 emit_pic_move (rtx *operands, enum machine_mode mode ATTRIBUTE_UNUSED)
1971 rtx temp = reload_in_progress ? operands[0] : gen_reg_rtx (Pmode);
1973 gcc_assert (!TARGET_FDPIC || !(reload_in_progress || reload_completed));
1974 if (GET_CODE (operands[0]) == MEM && SYMBOLIC_CONST (operands[1]))
1975 operands[1] = force_reg (SImode, operands[1]);
1977 operands[1] = legitimize_pic_address (operands[1], temp,
1978 TARGET_FDPIC ? OUR_FDPIC_REG
1979 : pic_offset_table_rtx);
1982 /* Expand a move operation in mode MODE. The operands are in OPERANDS.
1983 Returns true if no further code must be generated, false if the caller
1984 should generate an insn to move OPERANDS[1] to OPERANDS[0]. */
1987 expand_move (rtx *operands, enum machine_mode mode)
1989 rtx op = operands[1];
1990 if ((TARGET_ID_SHARED_LIBRARY || TARGET_FDPIC)
1991 && SYMBOLIC_CONST (op))
1992 emit_pic_move (operands, mode);
1993 else if (mode == SImode && GET_CODE (op) == CONST
1994 && GET_CODE (XEXP (op, 0)) == PLUS
1995 && GET_CODE (XEXP (XEXP (op, 0), 0)) == SYMBOL_REF
1996 && !bfin_legitimate_constant_p (op))
1998 rtx dest = operands[0];
2000 gcc_assert (!reload_in_progress && !reload_completed);
2002 op0 = force_reg (mode, XEXP (op, 0));
2004 if (!insn_data[CODE_FOR_addsi3].operand[2].predicate (op1, mode))
2005 op1 = force_reg (mode, op1);
2006 if (GET_CODE (dest) == MEM)
2007 dest = gen_reg_rtx (mode);
2008 emit_insn (gen_addsi3 (dest, op0, op1));
2009 if (dest == operands[0])
2013 /* Don't generate memory->memory or constant->memory moves, go through a
2015 else if ((reload_in_progress | reload_completed) == 0
2016 && GET_CODE (operands[0]) == MEM
2017 && GET_CODE (operands[1]) != REG)
2018 operands[1] = force_reg (mode, operands[1]);
2022 /* Split one or more DImode RTL references into pairs of SImode
2023 references. The RTL can be REG, offsettable MEM, integer constant, or
2024 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
2025 split and "num" is its length. lo_half and hi_half are output arrays
2026 that parallel "operands". */
2029 split_di (rtx operands[], int num, rtx lo_half[], rtx hi_half[])
2033 rtx op = operands[num];
2035 /* simplify_subreg refuse to split volatile memory addresses,
2036 but we still have to handle it. */
2037 if (GET_CODE (op) == MEM)
2039 lo_half[num] = adjust_address (op, SImode, 0);
2040 hi_half[num] = adjust_address (op, SImode, 4);
2044 lo_half[num] = simplify_gen_subreg (SImode, op,
2045 GET_MODE (op) == VOIDmode
2046 ? DImode : GET_MODE (op), 0);
2047 hi_half[num] = simplify_gen_subreg (SImode, op,
2048 GET_MODE (op) == VOIDmode
2049 ? DImode : GET_MODE (op), 4);
2055 bfin_longcall_p (rtx op, int call_cookie)
2057 gcc_assert (GET_CODE (op) == SYMBOL_REF);
2058 if (call_cookie & CALL_SHORT)
2060 if (call_cookie & CALL_LONG)
2062 if (TARGET_LONG_CALLS)
2067 /* Expand a call instruction. FNADDR is the call target, RETVAL the return value.
2068 COOKIE is a CONST_INT holding the call_cookie prepared init_cumulative_args.
2069 SIBCALL is nonzero if this is a sibling call. */
2072 bfin_expand_call (rtx retval, rtx fnaddr, rtx callarg1, rtx cookie, int sibcall)
2074 rtx use = NULL, call;
2075 rtx callee = XEXP (fnaddr, 0);
2076 int nelts = 2 + !!sibcall;
2078 rtx picreg = get_hard_reg_initial_val (SImode, FDPIC_REGNO);
2081 /* In an untyped call, we can get NULL for operand 2. */
2082 if (cookie == NULL_RTX)
2083 cookie = const0_rtx;
2085 /* Static functions and indirect calls don't need the pic register. */
2086 if (!TARGET_FDPIC && flag_pic
2087 && GET_CODE (callee) == SYMBOL_REF
2088 && !SYMBOL_REF_LOCAL_P (callee))
2089 use_reg (&use, pic_offset_table_rtx);
2093 int caller_has_l1_text, callee_has_l1_text;
2095 caller_has_l1_text = callee_has_l1_text = 0;
2097 if (lookup_attribute ("l1_text",
2098 DECL_ATTRIBUTES (cfun->decl)) != NULL_TREE)
2099 caller_has_l1_text = 1;
2101 if (GET_CODE (callee) == SYMBOL_REF
2102 && SYMBOL_REF_DECL (callee) && DECL_P (SYMBOL_REF_DECL (callee))
2105 DECL_ATTRIBUTES (SYMBOL_REF_DECL (callee))) != NULL_TREE)
2106 callee_has_l1_text = 1;
2108 if (GET_CODE (callee) != SYMBOL_REF
2109 || bfin_longcall_p (callee, INTVAL (cookie))
2110 || (GET_CODE (callee) == SYMBOL_REF
2111 && !SYMBOL_REF_LOCAL_P (callee)
2112 && TARGET_INLINE_PLT)
2113 || caller_has_l1_text != callee_has_l1_text
2114 || (caller_has_l1_text && callee_has_l1_text
2115 && (GET_CODE (callee) != SYMBOL_REF
2116 || !SYMBOL_REF_LOCAL_P (callee))))
2119 if (! address_operand (addr, Pmode))
2120 addr = force_reg (Pmode, addr);
2122 fnaddr = gen_reg_rtx (SImode);
2123 emit_insn (gen_load_funcdescsi (fnaddr, addr));
2124 fnaddr = gen_rtx_MEM (Pmode, fnaddr);
2126 picreg = gen_reg_rtx (SImode);
2127 emit_insn (gen_load_funcdescsi (picreg,
2128 plus_constant (addr, 4)));
2133 else if ((!register_no_elim_operand (callee, Pmode)
2134 && GET_CODE (callee) != SYMBOL_REF)
2135 || (GET_CODE (callee) == SYMBOL_REF
2136 && ((TARGET_ID_SHARED_LIBRARY && !TARGET_LEAF_ID_SHARED_LIBRARY)
2137 || bfin_longcall_p (callee, INTVAL (cookie)))))
2139 callee = copy_to_mode_reg (Pmode, callee);
2140 fnaddr = gen_rtx_MEM (Pmode, callee);
2142 call = gen_rtx_CALL (VOIDmode, fnaddr, callarg1);
2145 call = gen_rtx_SET (VOIDmode, retval, call);
2147 pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nelts));
2149 XVECEXP (pat, 0, n++) = call;
2151 XVECEXP (pat, 0, n++) = gen_rtx_USE (VOIDmode, picreg);
2152 XVECEXP (pat, 0, n++) = gen_rtx_USE (VOIDmode, cookie);
2154 XVECEXP (pat, 0, n++) = gen_rtx_RETURN (VOIDmode);
2155 call = emit_call_insn (pat);
2157 CALL_INSN_FUNCTION_USAGE (call) = use;
2160 /* Return 1 if hard register REGNO can hold a value of machine-mode MODE. */
2163 hard_regno_mode_ok (int regno, enum machine_mode mode)
2165 /* Allow only dregs to store value of mode HI or QI */
2166 enum reg_class class = REGNO_REG_CLASS (regno);
2171 if (mode == V2HImode)
2172 return D_REGNO_P (regno);
2173 if (class == CCREGS)
2174 return mode == BImode;
2175 if (mode == PDImode || mode == V2PDImode)
2176 return regno == REG_A0 || regno == REG_A1;
2178 /* Allow all normal 32-bit regs, except REG_M3, in case regclass ever comes
2179 up with a bad register class (such as ALL_REGS) for DImode. */
2181 return regno < REG_M3;
2184 && TEST_HARD_REG_BIT (reg_class_contents[PROLOGUE_REGS], regno))
2187 return TEST_HARD_REG_BIT (reg_class_contents[MOST_REGS], regno);
2190 /* Implements target hook vector_mode_supported_p. */
2193 bfin_vector_mode_supported_p (enum machine_mode mode)
2195 return mode == V2HImode;
2198 /* Return the cost of moving data from a register in class CLASS1 to
2199 one in class CLASS2. A cost of 2 is the default. */
2202 bfin_register_move_cost (enum machine_mode mode,
2203 enum reg_class class1, enum reg_class class2)
2205 /* These need secondary reloads, so they're more expensive. */
2206 if ((class1 == CCREGS && class2 != DREGS)
2207 || (class1 != DREGS && class2 == CCREGS))
2210 /* If optimizing for size, always prefer reg-reg over reg-memory moves. */
2214 /* There are some stalls involved when moving from a DREG to a different
2215 class reg, and using the value in one of the following instructions.
2216 Attempt to model this by slightly discouraging such moves. */
2217 if (class1 == DREGS && class2 != DREGS)
2220 if (GET_MODE_CLASS (mode) == MODE_INT)
2222 /* Discourage trying to use the accumulators. */
2223 if (TEST_HARD_REG_BIT (reg_class_contents[class1], REG_A0)
2224 || TEST_HARD_REG_BIT (reg_class_contents[class1], REG_A1)
2225 || TEST_HARD_REG_BIT (reg_class_contents[class2], REG_A0)
2226 || TEST_HARD_REG_BIT (reg_class_contents[class2], REG_A1))
2232 /* Return the cost of moving data of mode M between a
2233 register and memory. A value of 2 is the default; this cost is
2234 relative to those in `REGISTER_MOVE_COST'.
2236 ??? In theory L1 memory has single-cycle latency. We should add a switch
2237 that tells the compiler whether we expect to use only L1 memory for the
2238 program; it'll make the costs more accurate. */
2241 bfin_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
2242 enum reg_class class,
2243 int in ATTRIBUTE_UNUSED)
2245 /* Make memory accesses slightly more expensive than any register-register
2246 move. Also, penalize non-DP registers, since they need secondary
2247 reloads to load and store. */
2248 if (! reg_class_subset_p (class, DPREGS))
2254 /* Inform reload about cases where moving X with a mode MODE to a register in
2255 CLASS requires an extra scratch register. Return the class needed for the
2256 scratch register. */
2258 static enum reg_class
2259 bfin_secondary_reload (bool in_p, rtx x, enum reg_class class,
2260 enum machine_mode mode, secondary_reload_info *sri)
2262 /* If we have HImode or QImode, we can only use DREGS as secondary registers;
2263 in most other cases we can also use PREGS. */
2264 enum reg_class default_class = GET_MODE_SIZE (mode) >= 4 ? DPREGS : DREGS;
2265 enum reg_class x_class = NO_REGS;
2266 enum rtx_code code = GET_CODE (x);
2269 x = SUBREG_REG (x), code = GET_CODE (x);
2272 int regno = REGNO (x);
2273 if (regno >= FIRST_PSEUDO_REGISTER)
2274 regno = reg_renumber[regno];
2279 x_class = REGNO_REG_CLASS (regno);
2282 /* We can be asked to reload (plus (FP) (large_constant)) into a DREG.
2283 This happens as a side effect of register elimination, and we need
2284 a scratch register to do it. */
2285 if (fp_plus_const_operand (x, mode))
2287 rtx op2 = XEXP (x, 1);
2288 int large_constant_p = ! satisfies_constraint_Ks7 (op2);
2290 if (class == PREGS || class == PREGS_CLOBBERED)
2292 /* If destination is a DREG, we can do this without a scratch register
2293 if the constant is valid for an add instruction. */
2294 if ((class == DREGS || class == DPREGS)
2295 && ! large_constant_p)
2297 /* Reloading to anything other than a DREG? Use a PREG scratch
2299 sri->icode = CODE_FOR_reload_insi;
2303 /* Data can usually be moved freely between registers of most classes.
2304 AREGS are an exception; they can only move to or from another register
2305 in AREGS or one in DREGS. They can also be assigned the constant 0. */
2306 if (x_class == AREGS || x_class == EVEN_AREGS || x_class == ODD_AREGS)
2307 return (class == DREGS || class == AREGS || class == EVEN_AREGS
2308 || class == ODD_AREGS
2311 if (class == AREGS || class == EVEN_AREGS || class == ODD_AREGS)
2315 sri->icode = in_p ? CODE_FOR_reload_inpdi : CODE_FOR_reload_outpdi;
2319 if (x != const0_rtx && x_class != DREGS)
2327 /* CCREGS can only be moved from/to DREGS. */
2328 if (class == CCREGS && x_class != DREGS)
2330 if (x_class == CCREGS && class != DREGS)
2333 /* All registers other than AREGS can load arbitrary constants. The only
2334 case that remains is MEM. */
2336 if (! reg_class_subset_p (class, default_class))
2337 return default_class;
2342 /* Implement TARGET_HANDLE_OPTION. */
2345 bfin_handle_option (size_t code, const char *arg, int value)
2349 case OPT_mshared_library_id_:
2350 if (value > MAX_LIBRARY_ID)
2351 error ("-mshared-library-id=%s is not between 0 and %d",
2352 arg, MAX_LIBRARY_ID);
2353 bfin_lib_id_given = 1;
2362 while ((p = bfin_cpus[i].name) != NULL)
2364 if (strncmp (arg, p, strlen (p)) == 0)
2371 error ("-mcpu=%s is not valid", arg);
2375 bfin_cpu_type = bfin_cpus[i].type;
2377 q = arg + strlen (p);
2381 bfin_si_revision = bfin_cpus[i].si_revision;
2382 bfin_workarounds |= bfin_cpus[i].workarounds;
2384 else if (strcmp (q, "-none") == 0)
2385 bfin_si_revision = -1;
2386 else if (strcmp (q, "-any") == 0)
2388 bfin_si_revision = 0xffff;
2389 while (bfin_cpus[i].type == bfin_cpu_type)
2391 bfin_workarounds |= bfin_cpus[i].workarounds;
2397 unsigned int si_major, si_minor;
2400 rev_len = strlen (q);
2402 if (sscanf (q, "-%u.%u%n", &si_major, &si_minor, &n) != 2
2404 || si_major > 0xff || si_minor > 0xff)
2406 invalid_silicon_revision:
2407 error ("-mcpu=%s has invalid silicon revision", arg);
2411 bfin_si_revision = (si_major << 8) | si_minor;
2413 while (bfin_cpus[i].type == bfin_cpu_type
2414 && bfin_cpus[i].si_revision != bfin_si_revision)
2417 if (bfin_cpus[i].type != bfin_cpu_type)
2418 goto invalid_silicon_revision;
2420 bfin_workarounds |= bfin_cpus[i].workarounds;
2423 if (bfin_cpu_type == BFIN_CPU_BF561)
2424 warning (0, "bf561 support is incomplete yet.");
2434 static struct machine_function *
2435 bfin_init_machine_status (void)
2437 struct machine_function *f;
2439 f = ggc_alloc_cleared (sizeof (struct machine_function));
2444 /* Implement the macro OVERRIDE_OPTIONS. */
2447 override_options (void)
2449 /* If processor type is not specified, enable all workarounds. */
2450 if (bfin_cpu_type == BFIN_CPU_UNKNOWN)
2454 for (i = 0; bfin_cpus[i].name != NULL; i++)
2455 bfin_workarounds |= bfin_cpus[i].workarounds;
2457 bfin_si_revision = 0xffff;
2460 if (bfin_csync_anomaly == 1)
2461 bfin_workarounds |= WA_SPECULATIVE_SYNCS;
2462 else if (bfin_csync_anomaly == 0)
2463 bfin_workarounds &= ~WA_SPECULATIVE_SYNCS;
2465 if (bfin_specld_anomaly == 1)
2466 bfin_workarounds |= WA_SPECULATIVE_LOADS;
2467 else if (bfin_specld_anomaly == 0)
2468 bfin_workarounds &= ~WA_SPECULATIVE_LOADS;
2470 if (TARGET_OMIT_LEAF_FRAME_POINTER)
2471 flag_omit_frame_pointer = 1;
2473 /* Library identification */
2474 if (bfin_lib_id_given && ! TARGET_ID_SHARED_LIBRARY)
2475 error ("-mshared-library-id= specified without -mid-shared-library");
2477 if (stack_limit_rtx && TARGET_STACK_CHECK_L1)
2478 error ("Can't use multiple stack checking methods together.");
2480 if (TARGET_ID_SHARED_LIBRARY && TARGET_FDPIC)
2481 error ("ID shared libraries and FD-PIC mode can't be used together.");
2483 /* Don't allow the user to specify -mid-shared-library and -msep-data
2484 together, as it makes little sense from a user's point of view... */
2485 if (TARGET_SEP_DATA && TARGET_ID_SHARED_LIBRARY)
2486 error ("cannot specify both -msep-data and -mid-shared-library");
2487 /* ... internally, however, it's nearly the same. */
2488 if (TARGET_SEP_DATA)
2489 target_flags |= MASK_ID_SHARED_LIBRARY | MASK_LEAF_ID_SHARED_LIBRARY;
2491 if (TARGET_ID_SHARED_LIBRARY && flag_pic == 0)
2494 /* There is no single unaligned SI op for PIC code. Sometimes we
2495 need to use ".4byte" and sometimes we need to use ".picptr".
2496 See bfin_assemble_integer for details. */
2498 targetm.asm_out.unaligned_op.si = 0;
2500 /* Silently turn off flag_pic if not doing FDPIC or ID shared libraries,
2501 since we don't support it and it'll just break. */
2502 if (flag_pic && !TARGET_FDPIC && !TARGET_ID_SHARED_LIBRARY)
2505 if (TARGET_MULTICORE && bfin_cpu_type != BFIN_CPU_BF561)
2506 error ("-mmulticore can only be used with BF561");
2508 if (TARGET_COREA && !TARGET_MULTICORE)
2509 error ("-mcorea should be used with -mmulticore");
2511 if (TARGET_COREB && !TARGET_MULTICORE)
2512 error ("-mcoreb should be used with -mmulticore");
2514 if (TARGET_COREA && TARGET_COREB)
2515 error ("-mcorea and -mcoreb can't be used together");
2517 flag_schedule_insns = 0;
2519 /* Passes after sched2 can break the helpful TImode annotations that
2520 haifa-sched puts on every insn. Just do scheduling in reorg. */
2521 bfin_flag_schedule_insns2 = flag_schedule_insns_after_reload;
2522 flag_schedule_insns_after_reload = 0;
2524 init_machine_status = bfin_init_machine_status;
2527 /* Return the destination address of BRANCH.
2528 We need to use this instead of get_attr_length, because the
2529 cbranch_with_nops pattern conservatively sets its length to 6, and
2530 we still prefer to use shorter sequences. */
2533 branch_dest (rtx branch)
2537 rtx pat = PATTERN (branch);
2538 if (GET_CODE (pat) == PARALLEL)
2539 pat = XVECEXP (pat, 0, 0);
2540 dest = SET_SRC (pat);
2541 if (GET_CODE (dest) == IF_THEN_ELSE)
2542 dest = XEXP (dest, 1);
2543 dest = XEXP (dest, 0);
2544 dest_uid = INSN_UID (dest);
2545 return INSN_ADDRESSES (dest_uid);
2548 /* Return nonzero if INSN is annotated with a REG_BR_PROB note that indicates
2549 it's a branch that's predicted taken. */
2552 cbranch_predicted_taken_p (rtx insn)
2554 rtx x = find_reg_note (insn, REG_BR_PROB, 0);
2558 int pred_val = INTVAL (XEXP (x, 0));
2560 return pred_val >= REG_BR_PROB_BASE / 2;
2566 /* Templates for use by asm_conditional_branch. */
2568 static const char *ccbranch_templates[][3] = {
2569 { "if !cc jump %3;", "if cc jump 4 (bp); jump.s %3;", "if cc jump 6 (bp); jump.l %3;" },
2570 { "if cc jump %3;", "if !cc jump 4 (bp); jump.s %3;", "if !cc jump 6 (bp); jump.l %3;" },
2571 { "if !cc jump %3 (bp);", "if cc jump 4; jump.s %3;", "if cc jump 6; jump.l %3;" },
2572 { "if cc jump %3 (bp);", "if !cc jump 4; jump.s %3;", "if !cc jump 6; jump.l %3;" },
2575 /* Output INSN, which is a conditional branch instruction with operands
2578 We deal with the various forms of conditional branches that can be generated
2579 by bfin_reorg to prevent the hardware from doing speculative loads, by
2580 - emitting a sufficient number of nops, if N_NOPS is nonzero, or
2581 - always emitting the branch as predicted taken, if PREDICT_TAKEN is true.
2582 Either of these is only necessary if the branch is short, otherwise the
2583 template we use ends in an unconditional jump which flushes the pipeline
2587 asm_conditional_branch (rtx insn, rtx *operands, int n_nops, int predict_taken)
2589 int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn));
2590 /* Note : offset for instructions like if cc jmp; jump.[sl] offset
2591 is to be taken from start of if cc rather than jump.
2592 Range for jump.s is (-4094, 4096) instead of (-4096, 4094)
2594 int len = (offset >= -1024 && offset <= 1022 ? 0
2595 : offset >= -4094 && offset <= 4096 ? 1
2597 int bp = predict_taken && len == 0 ? 1 : cbranch_predicted_taken_p (insn);
2598 int idx = (bp << 1) | (GET_CODE (operands[0]) == EQ ? BRF : BRT);
2599 output_asm_insn (ccbranch_templates[idx][len], operands);
2600 gcc_assert (n_nops == 0 || !bp);
2602 while (n_nops-- > 0)
2603 output_asm_insn ("nop;", NULL);
2606 /* Emit rtl for a comparison operation CMP in mode MODE. Operands have been
2607 stored in bfin_compare_op0 and bfin_compare_op1 already. */
2610 bfin_gen_compare (rtx cmp, enum machine_mode mode ATTRIBUTE_UNUSED)
2612 enum rtx_code code1, code2;
2613 rtx op0 = bfin_compare_op0, op1 = bfin_compare_op1;
2614 rtx tem = bfin_cc_rtx;
2615 enum rtx_code code = GET_CODE (cmp);
2617 /* If we have a BImode input, then we already have a compare result, and
2618 do not need to emit another comparison. */
2619 if (GET_MODE (op0) == BImode)
2621 gcc_assert ((code == NE || code == EQ) && op1 == const0_rtx);
2622 tem = op0, code2 = code;
2627 /* bfin has these conditions */
2637 code1 = reverse_condition (code);
2641 emit_insn (gen_rtx_SET (BImode, tem,
2642 gen_rtx_fmt_ee (code1, BImode, op0, op1)));
2645 return gen_rtx_fmt_ee (code2, BImode, tem, CONST0_RTX (BImode));
2648 /* Return nonzero iff C has exactly one bit set if it is interpreted
2649 as a 32-bit constant. */
2652 log2constp (unsigned HOST_WIDE_INT c)
2655 return c != 0 && (c & (c-1)) == 0;
2658 /* Returns the number of consecutive least significant zeros in the binary
2659 representation of *V.
2660 We modify *V to contain the original value arithmetically shifted right by
2661 the number of zeroes. */
2664 shiftr_zero (HOST_WIDE_INT *v)
2666 unsigned HOST_WIDE_INT tmp = *v;
2667 unsigned HOST_WIDE_INT sgn;
2673 sgn = tmp & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1));
2674 while ((tmp & 0x1) == 0 && n <= 32)
2676 tmp = (tmp >> 1) | sgn;
2683 /* After reload, split the load of an immediate constant. OPERANDS are the
2684 operands of the movsi_insn pattern which we are splitting. We return
2685 nonzero if we emitted a sequence to load the constant, zero if we emitted
2686 nothing because we want to use the splitter's default sequence. */
2689 split_load_immediate (rtx operands[])
2691 HOST_WIDE_INT val = INTVAL (operands[1]);
2693 HOST_WIDE_INT shifted = val;
2694 HOST_WIDE_INT shifted_compl = ~val;
2695 int num_zero = shiftr_zero (&shifted);
2696 int num_compl_zero = shiftr_zero (&shifted_compl);
2697 unsigned int regno = REGNO (operands[0]);
2699 /* This case takes care of single-bit set/clear constants, which we could
2700 also implement with BITSET/BITCLR. */
2702 && shifted >= -32768 && shifted < 65536
2703 && (D_REGNO_P (regno)
2704 || (regno >= REG_P0 && regno <= REG_P7 && num_zero <= 2)))
2706 emit_insn (gen_movsi (operands[0], GEN_INT (shifted)));
2707 emit_insn (gen_ashlsi3 (operands[0], operands[0], GEN_INT (num_zero)));
2712 tmp |= -(tmp & 0x8000);
2714 /* If high word has one bit set or clear, try to use a bit operation. */
2715 if (D_REGNO_P (regno))
2717 if (log2constp (val & 0xFFFF0000))
2719 emit_insn (gen_movsi (operands[0], GEN_INT (val & 0xFFFF)));
2720 emit_insn (gen_iorsi3 (operands[0], operands[0], GEN_INT (val & 0xFFFF0000)));
2723 else if (log2constp (val | 0xFFFF) && (val & 0x8000) != 0)
2725 emit_insn (gen_movsi (operands[0], GEN_INT (tmp)));
2726 emit_insn (gen_andsi3 (operands[0], operands[0], GEN_INT (val | 0xFFFF)));
2730 if (D_REGNO_P (regno))
2732 if (tmp >= -64 && tmp <= 63)
2734 emit_insn (gen_movsi (operands[0], GEN_INT (tmp)));
2735 emit_insn (gen_movstricthi_high (operands[0], GEN_INT (val & -65536)));
2739 if ((val & 0xFFFF0000) == 0)
2741 emit_insn (gen_movsi (operands[0], const0_rtx));
2742 emit_insn (gen_movsi_low (operands[0], operands[0], operands[1]));
2746 if ((val & 0xFFFF0000) == 0xFFFF0000)
2748 emit_insn (gen_movsi (operands[0], constm1_rtx));
2749 emit_insn (gen_movsi_low (operands[0], operands[0], operands[1]));
2754 /* Need DREGs for the remaining case. */
2759 && num_compl_zero && shifted_compl >= -64 && shifted_compl <= 63)
2761 /* If optimizing for size, generate a sequence that has more instructions
2763 emit_insn (gen_movsi (operands[0], GEN_INT (shifted_compl)));
2764 emit_insn (gen_ashlsi3 (operands[0], operands[0],
2765 GEN_INT (num_compl_zero)));
2766 emit_insn (gen_one_cmplsi2 (operands[0], operands[0]));
2772 /* Return true if the legitimate memory address for a memory operand of mode
2773 MODE. Return false if not. */
2776 bfin_valid_add (enum machine_mode mode, HOST_WIDE_INT value)
2778 unsigned HOST_WIDE_INT v = value > 0 ? value : -value;
2779 int sz = GET_MODE_SIZE (mode);
2780 int shift = sz == 1 ? 0 : sz == 2 ? 1 : 2;
2781 /* The usual offsettable_memref machinery doesn't work so well for this
2782 port, so we deal with the problem here. */
2783 if (value > 0 && sz == 8)
2785 return (v & ~(0x7fff << shift)) == 0;
2789 bfin_valid_reg_p (unsigned int regno, int strict, enum machine_mode mode,
2790 enum rtx_code outer_code)
2793 return REGNO_OK_FOR_BASE_STRICT_P (regno, mode, outer_code, SCRATCH);
2795 return REGNO_OK_FOR_BASE_NONSTRICT_P (regno, mode, outer_code, SCRATCH);
2799 bfin_legitimate_address_p (enum machine_mode mode, rtx x, int strict)
2801 switch (GET_CODE (x)) {
2803 if (bfin_valid_reg_p (REGNO (x), strict, mode, MEM))
2807 if (REG_P (XEXP (x, 0))
2808 && bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, PLUS)
2809 && ((GET_CODE (XEXP (x, 1)) == UNSPEC && mode == SImode)
2810 || (GET_CODE (XEXP (x, 1)) == CONST_INT
2811 && bfin_valid_add (mode, INTVAL (XEXP (x, 1))))))
2816 if (LEGITIMATE_MODE_FOR_AUTOINC_P (mode)
2817 && REG_P (XEXP (x, 0))
2818 && bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, POST_INC))
2821 if (LEGITIMATE_MODE_FOR_AUTOINC_P (mode)
2822 && XEXP (x, 0) == stack_pointer_rtx
2823 && REG_P (XEXP (x, 0))
2824 && bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, PRE_DEC))
2833 /* Decide whether we can force certain constants to memory. If we
2834 decide we can't, the caller should be able to cope with it in
2838 bfin_cannot_force_const_mem (rtx x ATTRIBUTE_UNUSED)
2840 /* We have only one class of non-legitimate constants, and our movsi
2841 expander knows how to handle them. Dropping these constants into the
2842 data section would only shift the problem - we'd still get relocs
2843 outside the object, in the data section rather than the text section. */
2847 /* Ensure that for any constant of the form symbol + offset, the offset
2848 remains within the object. Any other constants are ok.
2849 This ensures that flat binaries never have to deal with relocations
2850 crossing section boundaries. */
2853 bfin_legitimate_constant_p (rtx x)
2856 HOST_WIDE_INT offset;
2858 if (GET_CODE (x) != CONST)
2862 gcc_assert (GET_CODE (x) == PLUS);
2866 if (GET_CODE (sym) != SYMBOL_REF
2867 || GET_CODE (x) != CONST_INT)
2869 offset = INTVAL (x);
2871 if (SYMBOL_REF_DECL (sym) == 0)
2874 || offset >= int_size_in_bytes (TREE_TYPE (SYMBOL_REF_DECL (sym))))
2881 bfin_rtx_costs (rtx x, int code, int outer_code, int *total)
2883 int cost2 = COSTS_N_INSNS (1);
2889 if (outer_code == SET || outer_code == PLUS)
2890 *total = satisfies_constraint_Ks7 (x) ? 0 : cost2;
2891 else if (outer_code == AND)
2892 *total = log2constp (~INTVAL (x)) ? 0 : cost2;
2893 else if (outer_code == LE || outer_code == LT || outer_code == EQ)
2894 *total = (INTVAL (x) >= -4 && INTVAL (x) <= 3) ? 0 : cost2;
2895 else if (outer_code == LEU || outer_code == LTU)
2896 *total = (INTVAL (x) >= 0 && INTVAL (x) <= 7) ? 0 : cost2;
2897 else if (outer_code == MULT)
2898 *total = (INTVAL (x) == 2 || INTVAL (x) == 4) ? 0 : cost2;
2899 else if (outer_code == ASHIFT && (INTVAL (x) == 1 || INTVAL (x) == 2))
2901 else if (outer_code == ASHIFT || outer_code == ASHIFTRT
2902 || outer_code == LSHIFTRT)
2903 *total = (INTVAL (x) >= 0 && INTVAL (x) <= 31) ? 0 : cost2;
2904 else if (outer_code == IOR || outer_code == XOR)
2905 *total = (INTVAL (x) & (INTVAL (x) - 1)) == 0 ? 0 : cost2;
2914 *total = COSTS_N_INSNS (2);
2920 if (GET_MODE (x) == SImode)
2922 if (GET_CODE (op0) == MULT
2923 && GET_CODE (XEXP (op0, 1)) == CONST_INT)
2925 HOST_WIDE_INT val = INTVAL (XEXP (op0, 1));
2926 if (val == 2 || val == 4)
2929 *total += rtx_cost (XEXP (op0, 0), outer_code);
2930 *total += rtx_cost (op1, outer_code);
2935 if (GET_CODE (op0) != REG
2936 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
2937 *total += rtx_cost (op0, SET);
2938 #if 0 /* We'd like to do this for accuracy, but it biases the loop optimizer
2939 towards creating too many induction variables. */
2940 if (!reg_or_7bit_operand (op1, SImode))
2941 *total += rtx_cost (op1, SET);
2944 else if (GET_MODE (x) == DImode)
2947 if (GET_CODE (op1) != CONST_INT
2948 || !satisfies_constraint_Ks7 (op1))
2949 *total += rtx_cost (op1, PLUS);
2950 if (GET_CODE (op0) != REG
2951 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
2952 *total += rtx_cost (op0, PLUS);
2957 if (GET_MODE (x) == DImode)
2966 if (GET_MODE (x) == DImode)
2973 if (GET_CODE (op0) != REG
2974 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
2975 *total += rtx_cost (op0, code);
2985 /* Handle special cases of IOR: rotates, ALIGN insns, movstricthi_high. */
2988 if ((GET_CODE (op0) == LSHIFTRT && GET_CODE (op1) == ASHIFT)
2989 || (GET_CODE (op0) == ASHIFT && GET_CODE (op1) == ZERO_EXTEND)
2990 || (GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT)
2991 || (GET_CODE (op0) == AND && GET_CODE (op1) == CONST_INT))
2998 if (GET_CODE (op0) != REG
2999 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
3000 *total += rtx_cost (op0, code);
3002 if (GET_MODE (x) == DImode)
3008 if (GET_MODE (x) != SImode)
3013 if (! rhs_andsi3_operand (XEXP (x, 1), SImode))
3014 *total += rtx_cost (XEXP (x, 1), code);
3018 if (! regorlog2_operand (XEXP (x, 1), SImode))
3019 *total += rtx_cost (XEXP (x, 1), code);
3026 if (outer_code == SET
3027 && XEXP (x, 1) == const1_rtx
3028 && GET_CODE (XEXP (x, 2)) == CONST_INT)
3044 if (GET_CODE (op0) == GET_CODE (op1)
3045 && (GET_CODE (op0) == ZERO_EXTEND
3046 || GET_CODE (op0) == SIGN_EXTEND))
3048 *total = COSTS_N_INSNS (1);
3049 op0 = XEXP (op0, 0);
3050 op1 = XEXP (op1, 0);
3052 else if (optimize_size)
3053 *total = COSTS_N_INSNS (1);
3055 *total = COSTS_N_INSNS (3);
3057 if (GET_CODE (op0) != REG
3058 && (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
3059 *total += rtx_cost (op0, MULT);
3060 if (GET_CODE (op1) != REG
3061 && (GET_CODE (op1) != SUBREG || GET_CODE (SUBREG_REG (op1)) != REG))
3062 *total += rtx_cost (op1, MULT);
3068 *total = COSTS_N_INSNS (32);
3073 if (outer_code == SET)
3082 /* Used for communication between {push,pop}_multiple_operation (which
3083 we use not only as a predicate) and the corresponding output functions. */
3084 static int first_preg_to_save, first_dreg_to_save;
3087 push_multiple_operation (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
3089 int lastdreg = 8, lastpreg = 6;
3092 first_preg_to_save = lastpreg;
3093 first_dreg_to_save = lastdreg;
3094 for (i = 1, group = 0; i < XVECLEN (op, 0) - 1; i++)
3096 rtx t = XVECEXP (op, 0, i);
3100 if (GET_CODE (t) != SET)
3104 dest = SET_DEST (t);
3105 if (GET_CODE (dest) != MEM || ! REG_P (src))
3107 dest = XEXP (dest, 0);
3108 if (GET_CODE (dest) != PLUS
3109 || ! REG_P (XEXP (dest, 0))
3110 || REGNO (XEXP (dest, 0)) != REG_SP
3111 || GET_CODE (XEXP (dest, 1)) != CONST_INT
3112 || INTVAL (XEXP (dest, 1)) != -i * 4)
3115 regno = REGNO (src);
3118 if (D_REGNO_P (regno))
3121 first_dreg_to_save = lastdreg = regno - REG_R0;
3123 else if (regno >= REG_P0 && regno <= REG_P7)
3126 first_preg_to_save = lastpreg = regno - REG_P0;
3136 if (regno >= REG_P0 && regno <= REG_P7)
3139 first_preg_to_save = lastpreg = regno - REG_P0;
3141 else if (regno != REG_R0 + lastdreg + 1)
3146 else if (group == 2)
3148 if (regno != REG_P0 + lastpreg + 1)
3157 pop_multiple_operation (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
3159 int lastdreg = 8, lastpreg = 6;
3162 for (i = 1, group = 0; i < XVECLEN (op, 0); i++)
3164 rtx t = XVECEXP (op, 0, i);
3168 if (GET_CODE (t) != SET)
3172 dest = SET_DEST (t);
3173 if (GET_CODE (src) != MEM || ! REG_P (dest))
3175 src = XEXP (src, 0);
3179 if (! REG_P (src) || REGNO (src) != REG_SP)
3182 else if (GET_CODE (src) != PLUS
3183 || ! REG_P (XEXP (src, 0))
3184 || REGNO (XEXP (src, 0)) != REG_SP
3185 || GET_CODE (XEXP (src, 1)) != CONST_INT
3186 || INTVAL (XEXP (src, 1)) != (i - 1) * 4)
3189 regno = REGNO (dest);
3192 if (regno == REG_R7)
3197 else if (regno != REG_P0 + lastpreg - 1)
3202 else if (group == 1)
3204 if (regno != REG_R0 + lastdreg - 1)
3210 first_dreg_to_save = lastdreg;
3211 first_preg_to_save = lastpreg;
3215 /* Emit assembly code for one multi-register push described by INSN, with
3216 operands in OPERANDS. */
3219 output_push_multiple (rtx insn, rtx *operands)
3224 /* Validate the insn again, and compute first_[dp]reg_to_save. */
3225 ok = push_multiple_operation (PATTERN (insn), VOIDmode);
3228 if (first_dreg_to_save == 8)
3229 sprintf (buf, "[--sp] = ( p5:%d );\n", first_preg_to_save);
3230 else if (first_preg_to_save == 6)
3231 sprintf (buf, "[--sp] = ( r7:%d );\n", first_dreg_to_save);
3233 sprintf (buf, "[--sp] = ( r7:%d, p5:%d );\n",
3234 first_dreg_to_save, first_preg_to_save);
3236 output_asm_insn (buf, operands);
3239 /* Emit assembly code for one multi-register pop described by INSN, with
3240 operands in OPERANDS. */
3243 output_pop_multiple (rtx insn, rtx *operands)
3248 /* Validate the insn again, and compute first_[dp]reg_to_save. */
3249 ok = pop_multiple_operation (PATTERN (insn), VOIDmode);
3252 if (first_dreg_to_save == 8)
3253 sprintf (buf, "( p5:%d ) = [sp++];\n", first_preg_to_save);
3254 else if (first_preg_to_save == 6)
3255 sprintf (buf, "( r7:%d ) = [sp++];\n", first_dreg_to_save);
3257 sprintf (buf, "( r7:%d, p5:%d ) = [sp++];\n",
3258 first_dreg_to_save, first_preg_to_save);
3260 output_asm_insn (buf, operands);
3263 /* Adjust DST and SRC by OFFSET bytes, and generate one move in mode MODE. */
3266 single_move_for_movmem (rtx dst, rtx src, enum machine_mode mode, HOST_WIDE_INT offset)
3268 rtx scratch = gen_reg_rtx (mode);
3271 srcmem = adjust_address_nv (src, mode, offset);
3272 dstmem = adjust_address_nv (dst, mode, offset);
3273 emit_move_insn (scratch, srcmem);
3274 emit_move_insn (dstmem, scratch);
3277 /* Expand a string move operation of COUNT_EXP bytes from SRC to DST, with
3278 alignment ALIGN_EXP. Return true if successful, false if we should fall
3279 back on a different method. */
3282 bfin_expand_movmem (rtx dst, rtx src, rtx count_exp, rtx align_exp)
3284 rtx srcreg, destreg, countreg;
3285 HOST_WIDE_INT align = 0;
3286 unsigned HOST_WIDE_INT count = 0;
3288 if (GET_CODE (align_exp) == CONST_INT)
3289 align = INTVAL (align_exp);
3290 if (GET_CODE (count_exp) == CONST_INT)
3292 count = INTVAL (count_exp);
3294 if (!TARGET_INLINE_ALL_STRINGOPS && count > 64)
3299 /* If optimizing for size, only do single copies inline. */
3302 if (count == 2 && align < 2)
3304 if (count == 4 && align < 4)
3306 if (count != 1 && count != 2 && count != 4)
3309 if (align < 2 && count != 1)
3312 destreg = copy_to_mode_reg (Pmode, XEXP (dst, 0));
3313 if (destreg != XEXP (dst, 0))
3314 dst = replace_equiv_address_nv (dst, destreg);
3315 srcreg = copy_to_mode_reg (Pmode, XEXP (src, 0));
3316 if (srcreg != XEXP (src, 0))
3317 src = replace_equiv_address_nv (src, srcreg);
3319 if (count != 0 && align >= 2)
3321 unsigned HOST_WIDE_INT offset = 0;
3325 if ((count & ~3) == 4)
3327 single_move_for_movmem (dst, src, SImode, offset);
3330 else if (count & ~3)
3332 HOST_WIDE_INT new_count = ((count >> 2) & 0x3fffffff) - 1;
3333 countreg = copy_to_mode_reg (Pmode, GEN_INT (new_count));
3335 emit_insn (gen_rep_movsi (destreg, srcreg, countreg, destreg, srcreg));
3339 single_move_for_movmem (dst, src, HImode, offset);
3345 if ((count & ~1) == 2)
3347 single_move_for_movmem (dst, src, HImode, offset);
3350 else if (count & ~1)
3352 HOST_WIDE_INT new_count = ((count >> 1) & 0x7fffffff) - 1;
3353 countreg = copy_to_mode_reg (Pmode, GEN_INT (new_count));
3355 emit_insn (gen_rep_movhi (destreg, srcreg, countreg, destreg, srcreg));
3360 single_move_for_movmem (dst, src, QImode, offset);
3367 /* Compute the alignment for a local variable.
3368 TYPE is the data type, and ALIGN is the alignment that
3369 the object would ordinarily have. The value of this macro is used
3370 instead of that alignment to align the object. */
3373 bfin_local_alignment (tree type, int align)
3375 /* Increasing alignment for (relatively) big types allows the builtin
3376 memcpy can use 32 bit loads/stores. */
3377 if (TYPE_SIZE (type)
3378 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
3379 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) > 8
3380 || TREE_INT_CST_HIGH (TYPE_SIZE (type))) && align < 32)
3385 /* Implement TARGET_SCHED_ISSUE_RATE. */
3388 bfin_issue_rate (void)
3394 bfin_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
3396 enum attr_type insn_type, dep_insn_type;
3397 int dep_insn_code_number;
3399 /* Anti and output dependencies have zero cost. */
3400 if (REG_NOTE_KIND (link) != 0)
3403 dep_insn_code_number = recog_memoized (dep_insn);
3405 /* If we can't recognize the insns, we can't really do anything. */
3406 if (dep_insn_code_number < 0 || recog_memoized (insn) < 0)
3409 insn_type = get_attr_type (insn);
3410 dep_insn_type = get_attr_type (dep_insn);
3412 if (dep_insn_type == TYPE_MOVE || dep_insn_type == TYPE_MCLD)
3414 rtx pat = PATTERN (dep_insn);
3415 if (GET_CODE (pat) == PARALLEL)
3416 pat = XVECEXP (pat, 0, 0);
3417 rtx dest = SET_DEST (pat);
3418 rtx src = SET_SRC (pat);
3419 if (! ADDRESS_REGNO_P (REGNO (dest))
3420 || ! (MEM_P (src) || D_REGNO_P (REGNO (src))))
3422 return cost + (dep_insn_type == TYPE_MOVE ? 4 : 3);
3429 /* Increment the counter for the number of loop instructions in the
3430 current function. */
3433 bfin_hardware_loop (void)
3435 cfun->machine->has_hardware_loops++;
3438 /* Maximum loop nesting depth. */
3439 #define MAX_LOOP_DEPTH 2
3441 /* Maximum size of a loop. */
3442 #define MAX_LOOP_LENGTH 2042
3444 /* Maximum distance of the LSETUP instruction from the loop start. */
3445 #define MAX_LSETUP_DISTANCE 30
3447 /* We need to keep a vector of loops */
3448 typedef struct loop_info *loop_info;
3449 DEF_VEC_P (loop_info);
3450 DEF_VEC_ALLOC_P (loop_info,heap);
3452 /* Information about a loop we have found (or are in the process of
3454 struct loop_info GTY (())
3456 /* loop number, for dumps */
3459 /* All edges that jump into and out of the loop. */
3460 VEC(edge,gc) *incoming;
3462 /* We can handle two cases: all incoming edges have the same destination
3463 block, or all incoming edges have the same source block. These two
3464 members are set to the common source or destination we found, or NULL
3465 if different blocks were found. If both are NULL the loop can't be
3467 basic_block incoming_src;
3468 basic_block incoming_dest;
3470 /* First block in the loop. This is the one branched to by the loop_end
3474 /* Last block in the loop (the one with the loop_end insn). */
3477 /* The successor block of the loop. This is the one the loop_end insn
3479 basic_block successor;
3481 /* The last instruction in the tail. */
3484 /* The loop_end insn. */
3487 /* The iteration register. */
3490 /* The new initialization insn. */
3493 /* The new initialization instruction. */
3496 /* The new label placed at the beginning of the loop. */
3499 /* The new label placed at the end of the loop. */
3502 /* The length of the loop. */
3505 /* The nesting depth of the loop. */
3508 /* Nonzero if we can't optimize this loop. */
3511 /* True if we have visited this loop. */
3514 /* True if this loop body clobbers any of LC0, LT0, or LB0. */
3517 /* True if this loop body clobbers any of LC1, LT1, or LB1. */
3520 /* Next loop in the graph. */
3521 struct loop_info *next;
3523 /* Immediate outer loop of this loop. */
3524 struct loop_info *outer;
3526 /* Vector of blocks only within the loop, including those within
3528 VEC (basic_block,heap) *blocks;
3530 /* Same information in a bitmap. */
3531 bitmap block_bitmap;
3533 /* Vector of inner loops within this loop */
3534 VEC (loop_info,heap) *loops;
3538 bfin_dump_loops (loop_info loops)
3542 for (loop = loops; loop; loop = loop->next)
3548 fprintf (dump_file, ";; loop %d: ", loop->loop_no);
3550 fprintf (dump_file, "(bad) ");
3551 fprintf (dump_file, "{head:%d, depth:%d}", loop->head->index, loop->depth);
3553 fprintf (dump_file, " blocks: [ ");
3554 for (ix = 0; VEC_iterate (basic_block, loop->blocks, ix, b); ix++)
3555 fprintf (dump_file, "%d ", b->index);
3556 fprintf (dump_file, "] ");
3558 fprintf (dump_file, " inner loops: [ ");
3559 for (ix = 0; VEC_iterate (loop_info, loop->loops, ix, i); ix++)
3560 fprintf (dump_file, "%d ", i->loop_no);
3561 fprintf (dump_file, "]\n");
3563 fprintf (dump_file, "\n");
3566 /* Scan the blocks of LOOP (and its inferiors) looking for basic block
3567 BB. Return true, if we find it. */
3570 bfin_bb_in_loop (loop_info loop, basic_block bb)
3572 return bitmap_bit_p (loop->block_bitmap, bb->index);
3575 /* Scan the blocks of LOOP (and its inferiors) looking for uses of
3576 REG. Return true, if we find any. Don't count the loop's loop_end
3577 insn if it matches LOOP_END. */
3580 bfin_scan_loop (loop_info loop, rtx reg, rtx loop_end)
3585 for (ix = 0; VEC_iterate (basic_block, loop->blocks, ix, bb); ix++)
3589 for (insn = BB_HEAD (bb);
3590 insn != NEXT_INSN (BB_END (bb));
3591 insn = NEXT_INSN (insn))
3595 if (insn == loop_end)
3597 if (reg_mentioned_p (reg, PATTERN (insn)))
3604 /* Estimate the length of INSN conservatively. */
3607 length_for_loop (rtx insn)
3610 if (JUMP_P (insn) && any_condjump_p (insn) && !optimize_size)
3612 if (ENABLE_WA_SPECULATIVE_SYNCS)
3614 else if (ENABLE_WA_SPECULATIVE_LOADS)
3617 else if (LABEL_P (insn))
3619 if (ENABLE_WA_SPECULATIVE_SYNCS)
3624 length += get_attr_length (insn);
3629 /* Optimize LOOP. */
3632 bfin_optimize_loop (loop_info loop)
3636 rtx insn, init_insn, last_insn, nop_insn;
3637 rtx loop_init, start_label, end_label;
3638 rtx reg_lc0, reg_lc1, reg_lt0, reg_lt1, reg_lb0, reg_lb1;
3640 rtx lc_reg, lt_reg, lb_reg;
3644 int inner_depth = 0;
3654 fprintf (dump_file, ";; loop %d bad when found\n", loop->loop_no);
3658 /* Every loop contains in its list of inner loops every loop nested inside
3659 it, even if there are intermediate loops. This works because we're doing
3660 a depth-first search here and never visit a loop more than once. */
3661 for (ix = 0; VEC_iterate (loop_info, loop->loops, ix, inner); ix++)
3663 bfin_optimize_loop (inner);
3665 if (!inner->bad && inner_depth < inner->depth)
3667 inner_depth = inner->depth;
3669 loop->clobber_loop0 |= inner->clobber_loop0;
3670 loop->clobber_loop1 |= inner->clobber_loop1;
3674 loop->depth = inner_depth + 1;
3675 if (loop->depth > MAX_LOOP_DEPTH)
3678 fprintf (dump_file, ";; loop %d too deep\n", loop->loop_no);
3682 /* Get the loop iteration register. */
3683 iter_reg = loop->iter_reg;
3685 if (!DPREG_P (iter_reg))
3688 fprintf (dump_file, ";; loop %d iteration count NOT in PREG or DREG\n",
3693 if (loop->incoming_src)
3695 /* Make sure the predecessor is before the loop start label, as required by
3696 the LSETUP instruction. */
3698 for (insn = BB_END (loop->incoming_src);
3699 insn && insn != loop->start_label;
3700 insn = NEXT_INSN (insn))
3701 length += length_for_loop (insn);
3706 fprintf (dump_file, ";; loop %d lsetup not before loop_start\n",
3711 if (length > MAX_LSETUP_DISTANCE)
3714 fprintf (dump_file, ";; loop %d lsetup too far away\n", loop->loop_no);
3719 /* Check if start_label appears before loop_end and calculate the
3720 offset between them. We calculate the length of instructions
3723 for (insn = loop->start_label;
3724 insn && insn != loop->loop_end;
3725 insn = NEXT_INSN (insn))
3726 length += length_for_loop (insn);
3731 fprintf (dump_file, ";; loop %d start_label not before loop_end\n",
3736 loop->length = length;
3737 if (loop->length > MAX_LOOP_LENGTH)
3740 fprintf (dump_file, ";; loop %d too long\n", loop->loop_no);
3744 /* Scan all the blocks to make sure they don't use iter_reg. */
3745 if (bfin_scan_loop (loop, iter_reg, loop->loop_end))
3748 fprintf (dump_file, ";; loop %d uses iterator\n", loop->loop_no);
3752 /* Scan all the insns to see if the loop body clobber
3753 any hardware loop registers. */
3755 reg_lc0 = gen_rtx_REG (SImode, REG_LC0);
3756 reg_lc1 = gen_rtx_REG (SImode, REG_LC1);
3757 reg_lt0 = gen_rtx_REG (SImode, REG_LT0);
3758 reg_lt1 = gen_rtx_REG (SImode, REG_LT1);
3759 reg_lb0 = gen_rtx_REG (SImode, REG_LB0);
3760 reg_lb1 = gen_rtx_REG (SImode, REG_LB1);
3762 for (ix = 0; VEC_iterate (basic_block, loop->blocks, ix, bb); ix++)
3766 for (insn = BB_HEAD (bb);
3767 insn != NEXT_INSN (BB_END (bb));
3768 insn = NEXT_INSN (insn))
3773 if (reg_set_p (reg_lc0, insn)
3774 || reg_set_p (reg_lt0, insn)
3775 || reg_set_p (reg_lb0, insn))
3776 loop->clobber_loop0 = 1;
3778 if (reg_set_p (reg_lc1, insn)
3779 || reg_set_p (reg_lt1, insn)
3780 || reg_set_p (reg_lb1, insn))
3781 loop->clobber_loop1 |= 1;
3785 if ((loop->clobber_loop0 && loop->clobber_loop1)
3786 || (loop->depth == MAX_LOOP_DEPTH && loop->clobber_loop0))
3788 loop->depth = MAX_LOOP_DEPTH + 1;
3790 fprintf (dump_file, ";; loop %d no loop reg available\n",
3795 /* There should be an instruction before the loop_end instruction
3796 in the same basic block. And the instruction must not be
3798 - CONDITIONAL BRANCH
3802 - Returns (RTS, RTN, etc.) */
3805 last_insn = PREV_INSN (loop->loop_end);
3809 for (; last_insn != PREV_INSN (BB_HEAD (bb));
3810 last_insn = PREV_INSN (last_insn))
3811 if (INSN_P (last_insn))
3814 if (last_insn != PREV_INSN (BB_HEAD (bb)))
3817 if (single_pred_p (bb)
3818 && single_pred (bb) != ENTRY_BLOCK_PTR)
3820 bb = single_pred (bb);
3821 last_insn = BB_END (bb);
3826 last_insn = NULL_RTX;
3834 fprintf (dump_file, ";; loop %d has no last instruction\n",
3839 if (JUMP_P (last_insn))
3841 loop_info inner = bb->aux;
3843 && inner->outer == loop
3844 && inner->loop_end == last_insn
3845 && inner->depth == 1)
3846 /* This jump_insn is the exact loop_end of an inner loop
3847 and to be optimized away. So use the inner's last_insn. */
3848 last_insn = inner->last_insn;
3852 fprintf (dump_file, ";; loop %d has bad last instruction\n",
3857 else if (CALL_P (last_insn)
3858 || (GET_CODE (PATTERN (last_insn)) != SEQUENCE
3859 && get_attr_type (last_insn) == TYPE_SYNC)
3860 || recog_memoized (last_insn) == CODE_FOR_return_internal)
3863 fprintf (dump_file, ";; loop %d has bad last instruction\n",
3868 if (GET_CODE (PATTERN (last_insn)) == ASM_INPUT
3869 || asm_noperands (PATTERN (last_insn)) >= 0
3870 || (GET_CODE (PATTERN (last_insn)) != SEQUENCE
3871 && get_attr_seq_insns (last_insn) == SEQ_INSNS_MULTI))
3873 nop_insn = emit_insn_after (gen_nop (), last_insn);
3874 last_insn = nop_insn;
3877 loop->last_insn = last_insn;
3879 /* The loop is good for replacement. */
3880 start_label = loop->start_label;
3881 end_label = gen_label_rtx ();
3882 iter_reg = loop->iter_reg;
3884 if (loop->depth == 1 && !loop->clobber_loop1)
3889 loop->clobber_loop1 = 1;
3896 loop->clobber_loop0 = 1;
3899 /* If iter_reg is a DREG, we need generate an instruction to load
3900 the loop count into LC register. */
3901 if (D_REGNO_P (REGNO (iter_reg)))
3903 init_insn = gen_movsi (lc_reg, iter_reg);
3904 loop_init = gen_lsetup_without_autoinit (lt_reg, start_label,
3908 else if (P_REGNO_P (REGNO (iter_reg)))
3910 init_insn = NULL_RTX;
3911 loop_init = gen_lsetup_with_autoinit (lt_reg, start_label,
3918 loop->init = init_insn;
3919 loop->end_label = end_label;
3920 loop->loop_init = loop_init;
3924 fprintf (dump_file, ";; replacing loop %d initializer with\n",
3926 print_rtl_single (dump_file, loop->loop_init);
3927 fprintf (dump_file, ";; replacing loop %d terminator with\n",
3929 print_rtl_single (dump_file, loop->loop_end);
3934 if (loop->init != NULL_RTX)
3935 emit_insn (loop->init);
3936 seq_end = emit_insn (loop->loop_init);
3941 if (loop->incoming_src)
3943 rtx prev = BB_END (loop->incoming_src);
3944 if (VEC_length (edge, loop->incoming) > 1
3945 || !(VEC_last (edge, loop->incoming)->flags & EDGE_FALLTHRU))
3947 gcc_assert (JUMP_P (prev));
3948 prev = PREV_INSN (prev);
3950 emit_insn_after (seq, prev);
3958 if (loop->head != loop->incoming_dest)
3960 FOR_EACH_EDGE (e, ei, loop->head->preds)
3962 if (e->flags & EDGE_FALLTHRU)
3964 rtx newjump = gen_jump (loop->start_label);
3965 emit_insn_before (newjump, BB_HEAD (loop->head));
3966 new_bb = create_basic_block (newjump, newjump, loop->head->prev_bb);
3967 gcc_assert (new_bb = loop->head->prev_bb);
3973 emit_insn_before (seq, BB_HEAD (loop->head));
3974 seq = emit_label_before (gen_label_rtx (), seq);
3976 new_bb = create_basic_block (seq, seq_end, loop->head->prev_bb);
3977 FOR_EACH_EDGE (e, ei, loop->incoming)
3979 if (!(e->flags & EDGE_FALLTHRU)
3980 || e->dest != loop->head)
3981 redirect_edge_and_branch_force (e, new_bb);
3983 redirect_edge_succ (e, new_bb);
3987 delete_insn (loop->loop_end);
3988 /* Insert the loop end label before the last instruction of the loop. */
3989 emit_label_before (loop->end_label, loop->last_insn);
3996 fprintf (dump_file, ";; loop %d is bad\n", loop->loop_no);
4000 if (DPREG_P (loop->iter_reg))
4002 /* If loop->iter_reg is a DREG or PREG, we can split it here
4003 without scratch register. */
4006 emit_insn_before (gen_addsi3 (loop->iter_reg,
4011 emit_insn_before (gen_cmpsi (loop->iter_reg, const0_rtx),
4014 insn = emit_jump_insn_before (gen_bne (loop->start_label),
4017 JUMP_LABEL (insn) = loop->start_label;
4018 LABEL_NUSES (loop->start_label)++;
4019 delete_insn (loop->loop_end);
4023 /* Called from bfin_reorg_loops when a potential loop end is found. LOOP is
4024 a newly set up structure describing the loop, it is this function's
4025 responsibility to fill most of it. TAIL_BB and TAIL_INSN point to the
4026 loop_end insn and its enclosing basic block. */
4029 bfin_discover_loop (loop_info loop, basic_block tail_bb, rtx tail_insn)
4033 VEC (basic_block,heap) *works = VEC_alloc (basic_block,heap,20);
4035 loop->tail = tail_bb;
4036 loop->head = BRANCH_EDGE (tail_bb)->dest;
4037 loop->successor = FALLTHRU_EDGE (tail_bb)->dest;
4038 loop->loop_end = tail_insn;
4039 loop->last_insn = NULL_RTX;
4040 loop->iter_reg = SET_DEST (XVECEXP (PATTERN (tail_insn), 0, 1));
4041 loop->depth = loop->length = 0;
4043 loop->clobber_loop0 = loop->clobber_loop1 = 0;
4046 loop->incoming = VEC_alloc (edge, gc, 2);
4047 loop->init = loop->loop_init = NULL_RTX;
4048 loop->start_label = XEXP (XEXP (SET_SRC (XVECEXP (PATTERN (tail_insn), 0, 0)), 1), 0);
4049 loop->end_label = NULL_RTX;
4052 VEC_safe_push (basic_block, heap, works, loop->head);
4054 while (VEC_iterate (basic_block, works, dwork++, bb))
4058 if (bb == EXIT_BLOCK_PTR)
4060 /* We've reached the exit block. The loop must be bad. */
4063 ";; Loop is bad - reached exit block while scanning\n");
4068 if (bitmap_bit_p (loop->block_bitmap, bb->index))
4071 /* We've not seen this block before. Add it to the loop's
4072 list and then add each successor to the work list. */
4074 VEC_safe_push (basic_block, heap, loop->blocks, bb);
4075 bitmap_set_bit (loop->block_bitmap, bb->index);
4079 FOR_EACH_EDGE (e, ei, bb->succs)
4081 basic_block succ = EDGE_SUCC (bb, ei.index)->dest;
4082 if (!REGNO_REG_SET_P (df_get_live_in (succ),
4083 REGNO (loop->iter_reg)))
4085 if (!VEC_space (basic_block, works, 1))
4089 VEC_block_remove (basic_block, works, 0, dwork);
4093 VEC_reserve (basic_block, heap, works, 1);
4095 VEC_quick_push (basic_block, works, succ);
4100 /* Find the predecessor, and make sure nothing else jumps into this loop. */
4104 for (dwork = 0; VEC_iterate (basic_block, loop->blocks, dwork, bb); dwork++)
4108 FOR_EACH_EDGE (e, ei, bb->preds)
4110 basic_block pred = e->src;
4112 if (!bfin_bb_in_loop (loop, pred))
4115 fprintf (dump_file, ";; Loop %d: incoming edge %d -> %d\n",
4116 loop->loop_no, pred->index,
4118 VEC_safe_push (edge, gc, loop->incoming, e);
4123 for (pass = 0, retry = 1; retry && pass < 2; pass++)
4130 FOR_EACH_EDGE (e, ei, loop->incoming)
4134 loop->incoming_src = e->src;
4135 loop->incoming_dest = e->dest;
4140 if (e->dest != loop->incoming_dest)
4141 loop->incoming_dest = NULL;
4142 if (e->src != loop->incoming_src)
4143 loop->incoming_src = NULL;
4145 if (loop->incoming_src == NULL && loop->incoming_dest == NULL)
4151 ";; retrying loop %d with forwarder blocks\n",
4159 ";; can't find suitable entry for loop %d\n",
4167 FOR_EACH_EDGE (e, ei, loop->incoming)
4169 if (forwarder_block_p (e->src))
4176 ";; Adding forwarder block %d to loop %d and retrying\n",
4177 e->src->index, loop->loop_no);
4178 VEC_safe_push (basic_block, heap, loop->blocks, e->src);
4179 bitmap_set_bit (loop->block_bitmap, e->src->index);
4180 FOR_EACH_EDGE (e2, ei2, e->src->preds)
4181 VEC_safe_push (edge, gc, loop->incoming, e2);
4182 VEC_unordered_remove (edge, loop->incoming, ei.index);
4192 VEC_free (basic_block, heap, works);
4195 /* Analyze the structure of the loops in the current function. Use STACK
4196 for bitmap allocations. Returns all the valid candidates for hardware
4197 loops found in this function. */
4199 bfin_discover_loops (bitmap_obstack *stack, FILE *dump_file)
4201 loop_info loops = NULL;
4207 /* Find all the possible loop tails. This means searching for every
4208 loop_end instruction. For each one found, create a loop_info
4209 structure and add the head block to the work list. */
4212 rtx tail = BB_END (bb);
4214 while (GET_CODE (tail) == NOTE)
4215 tail = PREV_INSN (tail);
4219 if (INSN_P (tail) && recog_memoized (tail) == CODE_FOR_loop_end)
4221 /* A possible loop end */
4223 loop = XNEW (struct loop_info);
4226 loop->loop_no = nloops++;
4227 loop->blocks = VEC_alloc (basic_block, heap, 20);
4228 loop->block_bitmap = BITMAP_ALLOC (stack);
4233 fprintf (dump_file, ";; potential loop %d ending at\n",
4235 print_rtl_single (dump_file, tail);
4238 bfin_discover_loop (loop, bb, tail);
4242 tmp_bitmap = BITMAP_ALLOC (stack);
4243 /* Compute loop nestings. */
4244 for (loop = loops; loop; loop = loop->next)
4250 for (other = loop->next; other; other = other->next)
4255 bitmap_and (tmp_bitmap, other->block_bitmap, loop->block_bitmap);
4256 if (bitmap_empty_p (tmp_bitmap))
4258 if (bitmap_equal_p (tmp_bitmap, other->block_bitmap))
4260 other->outer = loop;
4261 VEC_safe_push (loop_info, heap, loop->loops, other);
4263 else if (bitmap_equal_p (tmp_bitmap, loop->block_bitmap))
4265 loop->outer = other;
4266 VEC_safe_push (loop_info, heap, other->loops, loop);
4272 ";; can't find suitable nesting for loops %d and %d\n",
4273 loop->loop_no, other->loop_no);
4274 loop->bad = other->bad = 1;
4278 BITMAP_FREE (tmp_bitmap);
4283 /* Free up the loop structures in LOOPS. */
4285 free_loops (loop_info loops)
4289 loop_info loop = loops;
4291 VEC_free (loop_info, heap, loop->loops);
4292 VEC_free (basic_block, heap, loop->blocks);
4293 BITMAP_FREE (loop->block_bitmap);
4298 #define BB_AUX_INDEX(BB) ((unsigned)(BB)->aux)
4300 /* The taken-branch edge from the loop end can actually go forward. Since the
4301 Blackfin's LSETUP instruction requires that the loop end be after the loop
4302 start, try to reorder a loop's basic blocks when we find such a case. */
4304 bfin_reorder_loops (loop_info loops, FILE *dump_file)
4311 cfg_layout_initialize (0);
4313 for (loop = loops; loop; loop = loop->next)
4323 /* Recreate an index for basic blocks that represents their order. */
4324 for (bb = ENTRY_BLOCK_PTR->next_bb, index = 0;
4325 bb != EXIT_BLOCK_PTR;
4326 bb = bb->next_bb, index++)
4327 bb->aux = (PTR) index;
4329 if (BB_AUX_INDEX (loop->head) < BB_AUX_INDEX (loop->tail))
4332 FOR_EACH_EDGE (e, ei, loop->head->succs)
4334 if (bitmap_bit_p (loop->block_bitmap, e->dest->index)
4335 && BB_AUX_INDEX (e->dest) < BB_AUX_INDEX (loop->tail))
4337 basic_block start_bb = e->dest;
4338 basic_block start_prev_bb = start_bb->prev_bb;
4341 fprintf (dump_file, ";; Moving block %d before block %d\n",
4342 loop->head->index, start_bb->index);
4343 loop->head->prev_bb->next_bb = loop->head->next_bb;
4344 loop->head->next_bb->prev_bb = loop->head->prev_bb;
4346 loop->head->prev_bb = start_prev_bb;
4347 loop->head->next_bb = start_bb;
4348 start_prev_bb->next_bb = start_bb->prev_bb = loop->head;
4352 loops = loops->next;
4357 if (bb->next_bb != EXIT_BLOCK_PTR)
4358 bb->aux = bb->next_bb;
4362 cfg_layout_finalize ();
4366 /* Run from machine_dependent_reorg, this pass looks for doloop_end insns
4367 and tries to rewrite the RTL of these loops so that proper Blackfin
4368 hardware loops are generated. */
4371 bfin_reorg_loops (FILE *dump_file)
4373 loop_info loops = NULL;
4376 bitmap_obstack stack;
4378 bitmap_obstack_initialize (&stack);
4381 fprintf (dump_file, ";; Find loops, first pass\n\n");
4383 loops = bfin_discover_loops (&stack, dump_file);
4386 bfin_dump_loops (loops);
4388 bfin_reorder_loops (loops, dump_file);
4392 fprintf (dump_file, ";; Find loops, second pass\n\n");
4394 loops = bfin_discover_loops (&stack, dump_file);
4397 fprintf (dump_file, ";; All loops found:\n\n");
4398 bfin_dump_loops (loops);
4401 /* Now apply the optimizations. */
4402 for (loop = loops; loop; loop = loop->next)
4403 bfin_optimize_loop (loop);
4407 fprintf (dump_file, ";; After hardware loops optimization:\n\n");
4408 bfin_dump_loops (loops);
4414 print_rtl (dump_file, get_insns ());
4420 /* Possibly generate a SEQUENCE out of three insns found in SLOT.
4421 Returns true if we modified the insn chain, false otherwise. */
4423 gen_one_bundle (rtx slot[3])
4425 gcc_assert (slot[1] != NULL_RTX);
4427 /* Verify that we really can do the multi-issue. */
4430 rtx t = NEXT_INSN (slot[0]);
4431 while (t != slot[1])
4433 if (GET_CODE (t) != NOTE
4434 || NOTE_KIND (t) != NOTE_INSN_DELETED)
4441 rtx t = NEXT_INSN (slot[1]);
4442 while (t != slot[2])
4444 if (GET_CODE (t) != NOTE
4445 || NOTE_KIND (t) != NOTE_INSN_DELETED)
4451 if (slot[0] == NULL_RTX)
4453 slot[0] = emit_insn_before (gen_mnop (), slot[1]);
4454 df_insn_rescan (slot[0]);
4456 if (slot[2] == NULL_RTX)
4458 slot[2] = emit_insn_after (gen_forced_nop (), slot[1]);
4459 df_insn_rescan (slot[2]);
4462 /* Avoid line number information being printed inside one bundle. */
4463 if (INSN_LOCATOR (slot[1])
4464 && INSN_LOCATOR (slot[1]) != INSN_LOCATOR (slot[0]))
4465 INSN_LOCATOR (slot[1]) = INSN_LOCATOR (slot[0]);
4466 if (INSN_LOCATOR (slot[2])
4467 && INSN_LOCATOR (slot[2]) != INSN_LOCATOR (slot[0]))
4468 INSN_LOCATOR (slot[2]) = INSN_LOCATOR (slot[0]);
4470 /* Terminate them with "|| " instead of ";" in the output. */
4471 PUT_MODE (slot[0], SImode);
4472 PUT_MODE (slot[1], SImode);
4473 /* Terminate the bundle, for the benefit of reorder_var_tracking_notes. */
4474 PUT_MODE (slot[2], QImode);
4478 /* Go through all insns, and use the information generated during scheduling
4479 to generate SEQUENCEs to represent bundles of instructions issued
4483 bfin_gen_bundles (void)
4492 slot[0] = slot[1] = slot[2] = NULL_RTX;
4493 for (insn = BB_HEAD (bb);; insn = next)
4498 if (get_attr_type (insn) == TYPE_DSP32)
4500 else if (slot[1] == NULL_RTX)
4507 next = NEXT_INSN (insn);
4508 while (next && insn != BB_END (bb)
4510 && GET_CODE (PATTERN (next)) != USE
4511 && GET_CODE (PATTERN (next)) != CLOBBER))
4514 next = NEXT_INSN (insn);
4517 /* BB_END can change due to emitting extra NOPs, so check here. */
4518 at_end = insn == BB_END (bb);
4519 if (at_end || GET_MODE (next) == TImode)
4522 || !gen_one_bundle (slot))
4523 && slot[0] != NULL_RTX)
4525 rtx pat = PATTERN (slot[0]);
4526 if (GET_CODE (pat) == SET
4527 && GET_CODE (SET_SRC (pat)) == UNSPEC
4528 && XINT (SET_SRC (pat), 1) == UNSPEC_32BIT)
4530 SET_SRC (pat) = XVECEXP (SET_SRC (pat), 0, 0);
4531 INSN_CODE (slot[0]) = -1;
4532 df_insn_rescan (slot[0]);
4536 slot[0] = slot[1] = slot[2] = NULL_RTX;
4544 /* Ensure that no var tracking notes are emitted in the middle of a
4545 three-instruction bundle. */
4548 reorder_var_tracking_notes (void)
4554 rtx queue = NULL_RTX;
4555 bool in_bundle = false;
4557 for (insn = BB_HEAD (bb); insn != BB_END (bb); insn = next)
4559 next = NEXT_INSN (insn);
4563 /* Emit queued up notes at the last instruction of a bundle. */
4564 if (GET_MODE (insn) == QImode)
4568 rtx next_queue = PREV_INSN (queue);
4569 PREV_INSN (NEXT_INSN (insn)) = queue;
4570 NEXT_INSN (queue) = NEXT_INSN (insn);
4571 NEXT_INSN (insn) = queue;
4572 PREV_INSN (queue) = insn;
4577 else if (GET_MODE (insn) == SImode)
4580 else if (NOTE_P (insn) && NOTE_KIND (insn) == NOTE_INSN_VAR_LOCATION)
4584 rtx prev = PREV_INSN (insn);
4585 PREV_INSN (next) = prev;
4586 NEXT_INSN (prev) = next;
4588 PREV_INSN (insn) = queue;
4596 /* Return an insn type for INSN that can be used by the caller for anomaly
4597 workarounds. This differs from plain get_attr_type in that it handles
4600 static enum attr_type
4601 type_for_anomaly (rtx insn)
4603 rtx pat = PATTERN (insn);
4604 if (GET_CODE (pat) == SEQUENCE)
4607 t = get_attr_type (XVECEXP (pat, 0, 1));
4610 t = get_attr_type (XVECEXP (pat, 0, 2));
4616 return get_attr_type (insn);
4619 /* Return nonzero if INSN contains any loads that may trap. It handles
4620 SEQUENCEs correctly. */
4623 trapping_loads_p (rtx insn)
4625 rtx pat = PATTERN (insn);
4626 if (GET_CODE (pat) == SEQUENCE)
4629 t = get_attr_type (XVECEXP (pat, 0, 1));
4631 && may_trap_p (SET_SRC (PATTERN (XVECEXP (pat, 0, 1)))))
4633 t = get_attr_type (XVECEXP (pat, 0, 2));
4635 && may_trap_p (SET_SRC (PATTERN (XVECEXP (pat, 0, 2)))))
4640 return may_trap_p (SET_SRC (single_set (insn)));
4643 /* This function acts like NEXT_INSN, but is aware of three-insn bundles and
4644 skips all subsequent parallel instructions if INSN is the start of such
4647 find_next_insn_start (rtx insn)
4649 if (GET_MODE (insn) == SImode)
4651 while (GET_MODE (insn) != QImode)
4652 insn = NEXT_INSN (insn);
4654 return NEXT_INSN (insn);
4657 /* Return INSN if it is of TYPE_MCLD. Alternatively, if INSN is the start of
4658 a three-insn bundle, see if one of them is a load and return that if so.
4659 Return NULL_RTX if the insn does not contain loads. */
4661 find_load (rtx insn)
4663 if (get_attr_type (insn) == TYPE_MCLD)
4665 if (GET_MODE (insn) != SImode)
4668 insn = NEXT_INSN (insn);
4669 if ((GET_MODE (insn) == SImode || GET_MODE (insn) == QImode)
4670 && get_attr_type (insn) == TYPE_MCLD)
4672 } while (GET_MODE (insn) != QImode);
4676 /* We use the machine specific reorg pass for emitting CSYNC instructions
4677 after conditional branches as needed.
4679 The Blackfin is unusual in that a code sequence like
4682 may speculatively perform the load even if the condition isn't true. This
4683 happens for a branch that is predicted not taken, because the pipeline
4684 isn't flushed or stalled, so the early stages of the following instructions,
4685 which perform the memory reference, are allowed to execute before the
4686 jump condition is evaluated.
4687 Therefore, we must insert additional instructions in all places where this
4688 could lead to incorrect behavior. The manual recommends CSYNC, while
4689 VDSP seems to use NOPs (even though its corresponding compiler option is
4692 When optimizing for speed, we emit NOPs, which seems faster than a CSYNC.
4693 When optimizing for size, we turn the branch into a predicted taken one.
4694 This may be slower due to mispredicts, but saves code size. */
4700 rtx last_condjump = NULL_RTX;
4701 int cycles_since_jump = INT_MAX;
4703 /* We are freeing block_for_insn in the toplev to keep compatibility
4704 with old MDEP_REORGS that are not CFG based. Recompute it now. */
4705 compute_bb_for_insn ();
4707 if (bfin_flag_schedule_insns2)
4709 splitting_for_sched = 1;
4711 splitting_for_sched = 0;
4713 timevar_push (TV_SCHED2);
4715 timevar_pop (TV_SCHED2);
4717 /* Examine the schedule and insert nops as necessary for 64-bit parallel
4719 bfin_gen_bundles ();
4724 /* Doloop optimization */
4725 if (cfun->machine->has_hardware_loops)
4726 bfin_reorg_loops (dump_file);
4728 if (! ENABLE_WA_SPECULATIVE_LOADS && ! ENABLE_WA_SPECULATIVE_SYNCS)
4731 /* First pass: find predicted-false branches; if something after them
4732 needs nops, insert them or change the branch to predict true. */
4733 for (insn = get_insns (); insn; insn = next)
4737 next = find_next_insn_start (insn);
4739 if (NOTE_P (insn) || BARRIER_P (insn) || LABEL_P (insn))
4742 pat = PATTERN (insn);
4743 if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER
4744 || GET_CODE (pat) == ASM_INPUT || GET_CODE (pat) == ADDR_VEC
4745 || GET_CODE (pat) == ADDR_DIFF_VEC || asm_noperands (pat) >= 0)
4750 if (any_condjump_p (insn)
4751 && ! cbranch_predicted_taken_p (insn))
4753 last_condjump = insn;
4754 cycles_since_jump = 0;
4757 cycles_since_jump = INT_MAX;
4759 else if (INSN_P (insn))
4761 rtx load_insn = find_load (insn);
4762 enum attr_type type = type_for_anomaly (insn);
4763 int delay_needed = 0;
4764 if (cycles_since_jump < INT_MAX)
4765 cycles_since_jump++;
4767 if (load_insn && ENABLE_WA_SPECULATIVE_LOADS)
4769 if (trapping_loads_p (load_insn))
4772 else if (type == TYPE_SYNC && ENABLE_WA_SPECULATIVE_SYNCS)
4775 if (delay_needed > cycles_since_jump)
4779 rtx *op = recog_data.operand;
4781 delay_needed -= cycles_since_jump;
4783 extract_insn (last_condjump);
4786 pat = gen_cbranch_predicted_taken (op[0], op[1], op[2],
4788 cycles_since_jump = INT_MAX;
4791 /* Do not adjust cycles_since_jump in this case, so that
4792 we'll increase the number of NOPs for a subsequent insn
4794 pat = gen_cbranch_with_nops (op[0], op[1], op[2], op[3],
4795 GEN_INT (delay_needed));
4796 PATTERN (last_condjump) = pat;
4797 INSN_CODE (last_condjump) = recog (pat, insn, &num_clobbers);
4801 /* Second pass: for predicted-true branches, see if anything at the
4802 branch destination needs extra nops. */
4803 if (! ENABLE_WA_SPECULATIVE_SYNCS)
4806 if (! ENABLE_WA_RETS)
4809 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
4812 && any_condjump_p (insn)
4813 && (INSN_CODE (insn) == CODE_FOR_cbranch_predicted_taken
4814 || cbranch_predicted_taken_p (insn)))
4816 rtx target = JUMP_LABEL (insn);
4818 cycles_since_jump = 0;
4819 for (; target && cycles_since_jump < 3; target = NEXT_INSN (target))
4823 if (NOTE_P (target) || BARRIER_P (target) || LABEL_P (target))
4826 pat = PATTERN (target);
4827 if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER
4828 || GET_CODE (pat) == ASM_INPUT || GET_CODE (pat) == ADDR_VEC
4829 || GET_CODE (pat) == ADDR_DIFF_VEC || asm_noperands (pat) >= 0)
4832 if (INSN_P (target))
4834 enum attr_type type = type_for_anomaly (target);
4835 int delay_needed = 0;
4836 if (cycles_since_jump < INT_MAX)
4837 cycles_since_jump++;
4839 if (type == TYPE_SYNC && ENABLE_WA_SPECULATIVE_SYNCS)
4842 if (delay_needed > cycles_since_jump)
4844 rtx prev = prev_real_insn (label);
4845 delay_needed -= cycles_since_jump;
4847 fprintf (dump_file, "Adding %d nops after %d\n",
4848 delay_needed, INSN_UID (label));
4850 && INSN_CODE (prev) == CODE_FOR_cbranch_with_nops)
4857 "Reducing nops on insn %d.\n",
4860 x = XVECEXP (x, 0, 1);
4861 v = INTVAL (XVECEXP (x, 0, 0)) - delay_needed;
4862 XVECEXP (x, 0, 0) = GEN_INT (v);
4864 while (delay_needed-- > 0)
4865 emit_insn_after (gen_nop (), label);
4873 if (bfin_flag_var_tracking)
4875 timevar_push (TV_VAR_TRACKING);
4876 variable_tracking_main ();
4877 reorder_var_tracking_notes ();
4878 timevar_pop (TV_VAR_TRACKING);
4880 df_finish_pass (false);
4883 /* Handle interrupt_handler, exception_handler and nmi_handler function
4884 attributes; arguments as in struct attribute_spec.handler. */
4887 handle_int_attribute (tree *node, tree name,
4888 tree args ATTRIBUTE_UNUSED,
4889 int flags ATTRIBUTE_UNUSED,
4893 if (TREE_CODE (x) == FUNCTION_DECL)
4896 if (TREE_CODE (x) != FUNCTION_TYPE)
4898 warning (OPT_Wattributes, "%qs attribute only applies to functions",
4899 IDENTIFIER_POINTER (name));
4900 *no_add_attrs = true;
4902 else if (funkind (x) != SUBROUTINE)
4903 error ("multiple function type attributes specified");
4908 /* Return 0 if the attributes for two types are incompatible, 1 if they
4909 are compatible, and 2 if they are nearly compatible (which causes a
4910 warning to be generated). */
4913 bfin_comp_type_attributes (const_tree type1, const_tree type2)
4915 e_funkind kind1, kind2;
4917 if (TREE_CODE (type1) != FUNCTION_TYPE)
4920 kind1 = funkind (type1);
4921 kind2 = funkind (type2);
4926 /* Check for mismatched modifiers */
4927 if (!lookup_attribute ("nesting", TYPE_ATTRIBUTES (type1))
4928 != !lookup_attribute ("nesting", TYPE_ATTRIBUTES (type2)))
4931 if (!lookup_attribute ("saveall", TYPE_ATTRIBUTES (type1))
4932 != !lookup_attribute ("saveall", TYPE_ATTRIBUTES (type2)))
4935 if (!lookup_attribute ("kspisusp", TYPE_ATTRIBUTES (type1))
4936 != !lookup_attribute ("kspisusp", TYPE_ATTRIBUTES (type2)))
4939 if (!lookup_attribute ("longcall", TYPE_ATTRIBUTES (type1))
4940 != !lookup_attribute ("longcall", TYPE_ATTRIBUTES (type2)))
4946 /* Handle a "longcall" or "shortcall" attribute; arguments as in
4947 struct attribute_spec.handler. */
4950 bfin_handle_longcall_attribute (tree *node, tree name,
4951 tree args ATTRIBUTE_UNUSED,
4952 int flags ATTRIBUTE_UNUSED,
4955 if (TREE_CODE (*node) != FUNCTION_TYPE
4956 && TREE_CODE (*node) != FIELD_DECL
4957 && TREE_CODE (*node) != TYPE_DECL)
4959 warning (OPT_Wattributes, "`%s' attribute only applies to functions",
4960 IDENTIFIER_POINTER (name));
4961 *no_add_attrs = true;
4964 if ((strcmp (IDENTIFIER_POINTER (name), "longcall") == 0
4965 && lookup_attribute ("shortcall", TYPE_ATTRIBUTES (*node)))
4966 || (strcmp (IDENTIFIER_POINTER (name), "shortcall") == 0
4967 && lookup_attribute ("longcall", TYPE_ATTRIBUTES (*node))))
4969 warning (OPT_Wattributes,
4970 "can't apply both longcall and shortcall attributes to the same function");
4971 *no_add_attrs = true;
4977 /* Handle a "l1_text" attribute; arguments as in
4978 struct attribute_spec.handler. */
4981 bfin_handle_l1_text_attribute (tree *node, tree name, tree ARG_UNUSED (args),
4982 int ARG_UNUSED (flags), bool *no_add_attrs)
4986 if (TREE_CODE (decl) != FUNCTION_DECL)
4988 error ("`%s' attribute only applies to functions",
4989 IDENTIFIER_POINTER (name));
4990 *no_add_attrs = true;
4993 /* The decl may have already been given a section attribute
4994 from a previous declaration. Ensure they match. */
4995 else if (DECL_SECTION_NAME (decl) != NULL_TREE
4996 && strcmp (TREE_STRING_POINTER (DECL_SECTION_NAME (decl)),
4999 error ("section of %q+D conflicts with previous declaration",
5001 *no_add_attrs = true;
5004 DECL_SECTION_NAME (decl) = build_string (9, ".l1.text");
5009 /* Handle a "l1_data", "l1_data_A" or "l1_data_B" attribute;
5010 arguments as in struct attribute_spec.handler. */
5013 bfin_handle_l1_data_attribute (tree *node, tree name, tree ARG_UNUSED (args),
5014 int ARG_UNUSED (flags), bool *no_add_attrs)
5018 if (TREE_CODE (decl) != VAR_DECL)
5020 error ("`%s' attribute only applies to variables",
5021 IDENTIFIER_POINTER (name));
5022 *no_add_attrs = true;
5024 else if (current_function_decl != NULL_TREE
5025 && !TREE_STATIC (decl))
5027 error ("`%s' attribute cannot be specified for local variables",
5028 IDENTIFIER_POINTER (name));
5029 *no_add_attrs = true;
5033 const char *section_name;
5035 if (strcmp (IDENTIFIER_POINTER (name), "l1_data") == 0)
5036 section_name = ".l1.data";
5037 else if (strcmp (IDENTIFIER_POINTER (name), "l1_data_A") == 0)
5038 section_name = ".l1.data.A";
5039 else if (strcmp (IDENTIFIER_POINTER (name), "l1_data_B") == 0)
5040 section_name = ".l1.data.B";
5044 /* The decl may have already been given a section attribute
5045 from a previous declaration. Ensure they match. */
5046 if (DECL_SECTION_NAME (decl) != NULL_TREE
5047 && strcmp (TREE_STRING_POINTER (DECL_SECTION_NAME (decl)),
5050 error ("section of %q+D conflicts with previous declaration",
5052 *no_add_attrs = true;
5055 DECL_SECTION_NAME (decl)
5056 = build_string (strlen (section_name) + 1, section_name);
5062 /* Table of valid machine attributes. */
5063 const struct attribute_spec bfin_attribute_table[] =
5065 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
5066 { "interrupt_handler", 0, 0, false, true, true, handle_int_attribute },
5067 { "exception_handler", 0, 0, false, true, true, handle_int_attribute },
5068 { "nmi_handler", 0, 0, false, true, true, handle_int_attribute },
5069 { "nesting", 0, 0, false, true, true, NULL },
5070 { "kspisusp", 0, 0, false, true, true, NULL },
5071 { "saveall", 0, 0, false, true, true, NULL },
5072 { "longcall", 0, 0, false, true, true, bfin_handle_longcall_attribute },
5073 { "shortcall", 0, 0, false, true, true, bfin_handle_longcall_attribute },
5074 { "l1_text", 0, 0, true, false, false, bfin_handle_l1_text_attribute },
5075 { "l1_data", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
5076 { "l1_data_A", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
5077 { "l1_data_B", 0, 0, true, false, false, bfin_handle_l1_data_attribute },
5078 { NULL, 0, 0, false, false, false, NULL }
5081 /* Implementation of TARGET_ASM_INTEGER. When using FD-PIC, we need to
5082 tell the assembler to generate pointers to function descriptors in
5086 bfin_assemble_integer (rtx value, unsigned int size, int aligned_p)
5088 if (TARGET_FDPIC && size == UNITS_PER_WORD)
5090 if (GET_CODE (value) == SYMBOL_REF
5091 && SYMBOL_REF_FUNCTION_P (value))
5093 fputs ("\t.picptr\tfuncdesc(", asm_out_file);
5094 output_addr_const (asm_out_file, value);
5095 fputs (")\n", asm_out_file);
5100 /* We've set the unaligned SI op to NULL, so we always have to
5101 handle the unaligned case here. */
5102 assemble_integer_with_op ("\t.4byte\t", value);
5106 return default_assemble_integer (value, size, aligned_p);
5109 /* Output the assembler code for a thunk function. THUNK_DECL is the
5110 declaration for the thunk function itself, FUNCTION is the decl for
5111 the target function. DELTA is an immediate constant offset to be
5112 added to THIS. If VCALL_OFFSET is nonzero, the word at
5113 *(*this + vcall_offset) should be added to THIS. */
5116 bfin_output_mi_thunk (FILE *file ATTRIBUTE_UNUSED,
5117 tree thunk ATTRIBUTE_UNUSED, HOST_WIDE_INT delta,
5118 HOST_WIDE_INT vcall_offset, tree function)
5121 /* The this parameter is passed as the first argument. */
5122 rtx this = gen_rtx_REG (Pmode, REG_R0);
5124 /* Adjust the this parameter by a fixed constant. */
5128 if (delta >= -64 && delta <= 63)
5130 xops[0] = GEN_INT (delta);
5131 output_asm_insn ("%1 += %0;", xops);
5133 else if (delta >= -128 && delta < -64)
5135 xops[0] = GEN_INT (delta + 64);
5136 output_asm_insn ("%1 += -64; %1 += %0;", xops);
5138 else if (delta > 63 && delta <= 126)
5140 xops[0] = GEN_INT (delta - 63);
5141 output_asm_insn ("%1 += 63; %1 += %0;", xops);
5145 xops[0] = GEN_INT (delta);
5146 output_asm_insn ("r3.l = %h0; r3.h = %d0; %1 = %1 + r3;", xops);
5150 /* Adjust the this parameter by a value stored in the vtable. */
5153 rtx p2tmp = gen_rtx_REG (Pmode, REG_P2);
5154 rtx tmp = gen_rtx_REG (Pmode, REG_R3);
5158 output_asm_insn ("%2 = r0; %2 = [%2];", xops);
5160 /* Adjust the this parameter. */
5161 xops[0] = gen_rtx_MEM (Pmode, plus_constant (p2tmp, vcall_offset));
5162 if (!memory_operand (xops[0], Pmode))
5164 rtx tmp2 = gen_rtx_REG (Pmode, REG_P1);
5165 xops[0] = GEN_INT (vcall_offset);
5167 output_asm_insn ("%h1 = %h0; %d1 = %d0; %2 = %2 + %1", xops);
5168 xops[0] = gen_rtx_MEM (Pmode, p2tmp);
5171 output_asm_insn ("%1 = %0; %2 = %2 + %1;", xops);
5174 xops[0] = XEXP (DECL_RTL (function), 0);
5175 if (1 || !flag_pic || (*targetm.binds_local_p) (function))
5176 output_asm_insn ("jump.l\t%P0", xops);
5179 /* Codes for all the Blackfin builtins. */
5185 BFIN_BUILTIN_COMPOSE_2X16,
5186 BFIN_BUILTIN_EXTRACTLO,
5187 BFIN_BUILTIN_EXTRACTHI,
5189 BFIN_BUILTIN_SSADD_2X16,
5190 BFIN_BUILTIN_SSSUB_2X16,
5191 BFIN_BUILTIN_SSADDSUB_2X16,
5192 BFIN_BUILTIN_SSSUBADD_2X16,
5193 BFIN_BUILTIN_MULT_2X16,
5194 BFIN_BUILTIN_MULTR_2X16,
5195 BFIN_BUILTIN_NEG_2X16,
5196 BFIN_BUILTIN_ABS_2X16,
5197 BFIN_BUILTIN_MIN_2X16,
5198 BFIN_BUILTIN_MAX_2X16,
5200 BFIN_BUILTIN_SSADD_1X16,
5201 BFIN_BUILTIN_SSSUB_1X16,
5202 BFIN_BUILTIN_MULT_1X16,
5203 BFIN_BUILTIN_MULTR_1X16,
5204 BFIN_BUILTIN_NORM_1X16,
5205 BFIN_BUILTIN_NEG_1X16,
5206 BFIN_BUILTIN_ABS_1X16,
5207 BFIN_BUILTIN_MIN_1X16,
5208 BFIN_BUILTIN_MAX_1X16,
5210 BFIN_BUILTIN_SUM_2X16,
5211 BFIN_BUILTIN_DIFFHL_2X16,
5212 BFIN_BUILTIN_DIFFLH_2X16,
5214 BFIN_BUILTIN_SSADD_1X32,
5215 BFIN_BUILTIN_SSSUB_1X32,
5216 BFIN_BUILTIN_NORM_1X32,
5217 BFIN_BUILTIN_ROUND_1X32,
5218 BFIN_BUILTIN_NEG_1X32,
5219 BFIN_BUILTIN_ABS_1X32,
5220 BFIN_BUILTIN_MIN_1X32,
5221 BFIN_BUILTIN_MAX_1X32,
5222 BFIN_BUILTIN_MULT_1X32,
5223 BFIN_BUILTIN_MULT_1X32X32,
5224 BFIN_BUILTIN_MULT_1X32X32NS,
5226 BFIN_BUILTIN_MULHISILL,
5227 BFIN_BUILTIN_MULHISILH,
5228 BFIN_BUILTIN_MULHISIHL,
5229 BFIN_BUILTIN_MULHISIHH,
5231 BFIN_BUILTIN_LSHIFT_1X16,
5232 BFIN_BUILTIN_LSHIFT_2X16,
5233 BFIN_BUILTIN_SSASHIFT_1X16,
5234 BFIN_BUILTIN_SSASHIFT_2X16,
5235 BFIN_BUILTIN_SSASHIFT_1X32,
5237 BFIN_BUILTIN_CPLX_MUL_16,
5238 BFIN_BUILTIN_CPLX_MAC_16,
5239 BFIN_BUILTIN_CPLX_MSU_16,
5241 BFIN_BUILTIN_CPLX_MUL_16_S40,
5242 BFIN_BUILTIN_CPLX_MAC_16_S40,
5243 BFIN_BUILTIN_CPLX_MSU_16_S40,
5245 BFIN_BUILTIN_CPLX_SQU,
5247 BFIN_BUILTIN_LOADBYTES,
5252 #define def_builtin(NAME, TYPE, CODE) \
5254 add_builtin_function ((NAME), (TYPE), (CODE), BUILT_IN_MD, \
5258 /* Set up all builtin functions for this target. */
5260 bfin_init_builtins (void)
5262 tree V2HI_type_node = build_vector_type_for_mode (intHI_type_node, V2HImode);
5263 tree void_ftype_void
5264 = build_function_type (void_type_node, void_list_node);
5265 tree short_ftype_short
5266 = build_function_type_list (short_integer_type_node, short_integer_type_node,
5268 tree short_ftype_int_int
5269 = build_function_type_list (short_integer_type_node, integer_type_node,
5270 integer_type_node, NULL_TREE);
5271 tree int_ftype_int_int
5272 = build_function_type_list (integer_type_node, integer_type_node,
5273 integer_type_node, NULL_TREE);
5275 = build_function_type_list (integer_type_node, integer_type_node,
5277 tree short_ftype_int
5278 = build_function_type_list (short_integer_type_node, integer_type_node,
5280 tree int_ftype_v2hi_v2hi
5281 = build_function_type_list (integer_type_node, V2HI_type_node,
5282 V2HI_type_node, NULL_TREE);
5283 tree v2hi_ftype_v2hi_v2hi
5284 = build_function_type_list (V2HI_type_node, V2HI_type_node,
5285 V2HI_type_node, NULL_TREE);
5286 tree v2hi_ftype_v2hi_v2hi_v2hi
5287 = build_function_type_list (V2HI_type_node, V2HI_type_node,
5288 V2HI_type_node, V2HI_type_node, NULL_TREE);
5289 tree v2hi_ftype_int_int
5290 = build_function_type_list (V2HI_type_node, integer_type_node,
5291 integer_type_node, NULL_TREE);
5292 tree v2hi_ftype_v2hi_int
5293 = build_function_type_list (V2HI_type_node, V2HI_type_node,
5294 integer_type_node, NULL_TREE);
5295 tree int_ftype_short_short
5296 = build_function_type_list (integer_type_node, short_integer_type_node,
5297 short_integer_type_node, NULL_TREE);
5298 tree v2hi_ftype_v2hi
5299 = build_function_type_list (V2HI_type_node, V2HI_type_node, NULL_TREE);
5300 tree short_ftype_v2hi
5301 = build_function_type_list (short_integer_type_node, V2HI_type_node,
5304 = build_function_type_list (integer_type_node,
5305 build_pointer_type (integer_type_node),
5308 /* Add the remaining MMX insns with somewhat more complicated types. */
5309 def_builtin ("__builtin_bfin_csync", void_ftype_void, BFIN_BUILTIN_CSYNC);
5310 def_builtin ("__builtin_bfin_ssync", void_ftype_void, BFIN_BUILTIN_SSYNC);
5312 def_builtin ("__builtin_bfin_ones", short_ftype_int, BFIN_BUILTIN_ONES);
5314 def_builtin ("__builtin_bfin_compose_2x16", v2hi_ftype_int_int,
5315 BFIN_BUILTIN_COMPOSE_2X16);
5316 def_builtin ("__builtin_bfin_extract_hi", short_ftype_v2hi,
5317 BFIN_BUILTIN_EXTRACTHI);
5318 def_builtin ("__builtin_bfin_extract_lo", short_ftype_v2hi,
5319 BFIN_BUILTIN_EXTRACTLO);
5321 def_builtin ("__builtin_bfin_min_fr2x16", v2hi_ftype_v2hi_v2hi,
5322 BFIN_BUILTIN_MIN_2X16);
5323 def_builtin ("__builtin_bfin_max_fr2x16", v2hi_ftype_v2hi_v2hi,
5324 BFIN_BUILTIN_MAX_2X16);
5326 def_builtin ("__builtin_bfin_add_fr2x16", v2hi_ftype_v2hi_v2hi,
5327 BFIN_BUILTIN_SSADD_2X16);
5328 def_builtin ("__builtin_bfin_sub_fr2x16", v2hi_ftype_v2hi_v2hi,
5329 BFIN_BUILTIN_SSSUB_2X16);
5330 def_builtin ("__builtin_bfin_dspaddsubsat", v2hi_ftype_v2hi_v2hi,
5331 BFIN_BUILTIN_SSADDSUB_2X16);
5332 def_builtin ("__builtin_bfin_dspsubaddsat", v2hi_ftype_v2hi_v2hi,
5333 BFIN_BUILTIN_SSSUBADD_2X16);
5334 def_builtin ("__builtin_bfin_mult_fr2x16", v2hi_ftype_v2hi_v2hi,
5335 BFIN_BUILTIN_MULT_2X16);
5336 def_builtin ("__builtin_bfin_multr_fr2x16", v2hi_ftype_v2hi_v2hi,
5337 BFIN_BUILTIN_MULTR_2X16);
5338 def_builtin ("__builtin_bfin_negate_fr2x16", v2hi_ftype_v2hi,
5339 BFIN_BUILTIN_NEG_2X16);
5340 def_builtin ("__builtin_bfin_abs_fr2x16", v2hi_ftype_v2hi,
5341 BFIN_BUILTIN_ABS_2X16);
5343 def_builtin ("__builtin_bfin_min_fr1x16", short_ftype_int_int,
5344 BFIN_BUILTIN_MIN_1X16);
5345 def_builtin ("__builtin_bfin_max_fr1x16", short_ftype_int_int,
5346 BFIN_BUILTIN_MAX_1X16);
5348 def_builtin ("__builtin_bfin_add_fr1x16", short_ftype_int_int,
5349 BFIN_BUILTIN_SSADD_1X16);
5350 def_builtin ("__builtin_bfin_sub_fr1x16", short_ftype_int_int,
5351 BFIN_BUILTIN_SSSUB_1X16);
5352 def_builtin ("__builtin_bfin_mult_fr1x16", short_ftype_int_int,
5353 BFIN_BUILTIN_MULT_1X16);
5354 def_builtin ("__builtin_bfin_multr_fr1x16", short_ftype_int_int,
5355 BFIN_BUILTIN_MULTR_1X16);
5356 def_builtin ("__builtin_bfin_negate_fr1x16", short_ftype_short,
5357 BFIN_BUILTIN_NEG_1X16);
5358 def_builtin ("__builtin_bfin_abs_fr1x16", short_ftype_short,
5359 BFIN_BUILTIN_ABS_1X16);
5360 def_builtin ("__builtin_bfin_norm_fr1x16", short_ftype_int,
5361 BFIN_BUILTIN_NORM_1X16);
5363 def_builtin ("__builtin_bfin_sum_fr2x16", short_ftype_v2hi,
5364 BFIN_BUILTIN_SUM_2X16);
5365 def_builtin ("__builtin_bfin_diff_hl_fr2x16", short_ftype_v2hi,
5366 BFIN_BUILTIN_DIFFHL_2X16);
5367 def_builtin ("__builtin_bfin_diff_lh_fr2x16", short_ftype_v2hi,
5368 BFIN_BUILTIN_DIFFLH_2X16);
5370 def_builtin ("__builtin_bfin_mulhisill", int_ftype_v2hi_v2hi,
5371 BFIN_BUILTIN_MULHISILL);
5372 def_builtin ("__builtin_bfin_mulhisihl", int_ftype_v2hi_v2hi,
5373 BFIN_BUILTIN_MULHISIHL);
5374 def_builtin ("__builtin_bfin_mulhisilh", int_ftype_v2hi_v2hi,
5375 BFIN_BUILTIN_MULHISILH);
5376 def_builtin ("__builtin_bfin_mulhisihh", int_ftype_v2hi_v2hi,
5377 BFIN_BUILTIN_MULHISIHH);
5379 def_builtin ("__builtin_bfin_min_fr1x32", int_ftype_int_int,
5380 BFIN_BUILTIN_MIN_1X32);
5381 def_builtin ("__builtin_bfin_max_fr1x32", int_ftype_int_int,
5382 BFIN_BUILTIN_MAX_1X32);
5384 def_builtin ("__builtin_bfin_add_fr1x32", int_ftype_int_int,
5385 BFIN_BUILTIN_SSADD_1X32);
5386 def_builtin ("__builtin_bfin_sub_fr1x32", int_ftype_int_int,
5387 BFIN_BUILTIN_SSSUB_1X32);
5388 def_builtin ("__builtin_bfin_negate_fr1x32", int_ftype_int,
5389 BFIN_BUILTIN_NEG_1X32);
5390 def_builtin ("__builtin_bfin_abs_fr1x32", int_ftype_int,
5391 BFIN_BUILTIN_ABS_1X32);
5392 def_builtin ("__builtin_bfin_norm_fr1x32", short_ftype_int,
5393 BFIN_BUILTIN_NORM_1X32);
5394 def_builtin ("__builtin_bfin_round_fr1x32", short_ftype_int,
5395 BFIN_BUILTIN_ROUND_1X32);
5396 def_builtin ("__builtin_bfin_mult_fr1x32", int_ftype_short_short,
5397 BFIN_BUILTIN_MULT_1X32);
5398 def_builtin ("__builtin_bfin_mult_fr1x32x32", int_ftype_int_int,
5399 BFIN_BUILTIN_MULT_1X32X32);
5400 def_builtin ("__builtin_bfin_mult_fr1x32x32NS", int_ftype_int_int,
5401 BFIN_BUILTIN_MULT_1X32X32NS);
5404 def_builtin ("__builtin_bfin_shl_fr1x16", short_ftype_int_int,
5405 BFIN_BUILTIN_SSASHIFT_1X16);
5406 def_builtin ("__builtin_bfin_shl_fr2x16", v2hi_ftype_v2hi_int,
5407 BFIN_BUILTIN_SSASHIFT_2X16);
5408 def_builtin ("__builtin_bfin_lshl_fr1x16", short_ftype_int_int,
5409 BFIN_BUILTIN_LSHIFT_1X16);
5410 def_builtin ("__builtin_bfin_lshl_fr2x16", v2hi_ftype_v2hi_int,
5411 BFIN_BUILTIN_LSHIFT_2X16);
5412 def_builtin ("__builtin_bfin_shl_fr1x32", int_ftype_int_int,
5413 BFIN_BUILTIN_SSASHIFT_1X32);
5415 /* Complex numbers. */
5416 def_builtin ("__builtin_bfin_cmplx_add", v2hi_ftype_v2hi_v2hi,
5417 BFIN_BUILTIN_SSADD_2X16);
5418 def_builtin ("__builtin_bfin_cmplx_sub", v2hi_ftype_v2hi_v2hi,
5419 BFIN_BUILTIN_SSSUB_2X16);
5420 def_builtin ("__builtin_bfin_cmplx_mul", v2hi_ftype_v2hi_v2hi,
5421 BFIN_BUILTIN_CPLX_MUL_16);
5422 def_builtin ("__builtin_bfin_cmplx_mac", v2hi_ftype_v2hi_v2hi_v2hi,
5423 BFIN_BUILTIN_CPLX_MAC_16);
5424 def_builtin ("__builtin_bfin_cmplx_msu", v2hi_ftype_v2hi_v2hi_v2hi,
5425 BFIN_BUILTIN_CPLX_MSU_16);
5426 def_builtin ("__builtin_bfin_cmplx_mul_s40", v2hi_ftype_v2hi_v2hi,
5427 BFIN_BUILTIN_CPLX_MUL_16_S40);
5428 def_builtin ("__builtin_bfin_cmplx_mac_s40", v2hi_ftype_v2hi_v2hi_v2hi,
5429 BFIN_BUILTIN_CPLX_MAC_16_S40);
5430 def_builtin ("__builtin_bfin_cmplx_msu_s40", v2hi_ftype_v2hi_v2hi_v2hi,
5431 BFIN_BUILTIN_CPLX_MSU_16_S40);
5432 def_builtin ("__builtin_bfin_csqu_fr16", v2hi_ftype_v2hi,
5433 BFIN_BUILTIN_CPLX_SQU);
5435 /* "Unaligned" load. */
5436 def_builtin ("__builtin_bfin_loadbytes", int_ftype_pint,
5437 BFIN_BUILTIN_LOADBYTES);
5442 struct builtin_description
5444 const enum insn_code icode;
5445 const char *const name;
5446 const enum bfin_builtins code;
5450 static const struct builtin_description bdesc_2arg[] =
5452 { CODE_FOR_composev2hi, "__builtin_bfin_compose_2x16", BFIN_BUILTIN_COMPOSE_2X16, -1 },
5454 { CODE_FOR_ssashiftv2hi3, "__builtin_bfin_shl_fr2x16", BFIN_BUILTIN_SSASHIFT_2X16, -1 },
5455 { CODE_FOR_ssashifthi3, "__builtin_bfin_shl_fr1x16", BFIN_BUILTIN_SSASHIFT_1X16, -1 },
5456 { CODE_FOR_lshiftv2hi3, "__builtin_bfin_lshl_fr2x16", BFIN_BUILTIN_LSHIFT_2X16, -1 },
5457 { CODE_FOR_lshifthi3, "__builtin_bfin_lshl_fr1x16", BFIN_BUILTIN_LSHIFT_1X16, -1 },
5458 { CODE_FOR_ssashiftsi3, "__builtin_bfin_shl_fr1x32", BFIN_BUILTIN_SSASHIFT_1X32, -1 },
5460 { CODE_FOR_sminhi3, "__builtin_bfin_min_fr1x16", BFIN_BUILTIN_MIN_1X16, -1 },
5461 { CODE_FOR_smaxhi3, "__builtin_bfin_max_fr1x16", BFIN_BUILTIN_MAX_1X16, -1 },
5462 { CODE_FOR_ssaddhi3, "__builtin_bfin_add_fr1x16", BFIN_BUILTIN_SSADD_1X16, -1 },
5463 { CODE_FOR_sssubhi3, "__builtin_bfin_sub_fr1x16", BFIN_BUILTIN_SSSUB_1X16, -1 },
5465 { CODE_FOR_sminsi3, "__builtin_bfin_min_fr1x32", BFIN_BUILTIN_MIN_1X32, -1 },
5466 { CODE_FOR_smaxsi3, "__builtin_bfin_max_fr1x32", BFIN_BUILTIN_MAX_1X32, -1 },
5467 { CODE_FOR_ssaddsi3, "__builtin_bfin_add_fr1x32", BFIN_BUILTIN_SSADD_1X32, -1 },
5468 { CODE_FOR_sssubsi3, "__builtin_bfin_sub_fr1x32", BFIN_BUILTIN_SSSUB_1X32, -1 },
5470 { CODE_FOR_sminv2hi3, "__builtin_bfin_min_fr2x16", BFIN_BUILTIN_MIN_2X16, -1 },
5471 { CODE_FOR_smaxv2hi3, "__builtin_bfin_max_fr2x16", BFIN_BUILTIN_MAX_2X16, -1 },
5472 { CODE_FOR_ssaddv2hi3, "__builtin_bfin_add_fr2x16", BFIN_BUILTIN_SSADD_2X16, -1 },
5473 { CODE_FOR_sssubv2hi3, "__builtin_bfin_sub_fr2x16", BFIN_BUILTIN_SSSUB_2X16, -1 },
5474 { CODE_FOR_ssaddsubv2hi3, "__builtin_bfin_dspaddsubsat", BFIN_BUILTIN_SSADDSUB_2X16, -1 },
5475 { CODE_FOR_sssubaddv2hi3, "__builtin_bfin_dspsubaddsat", BFIN_BUILTIN_SSSUBADD_2X16, -1 },
5477 { CODE_FOR_flag_mulhisi, "__builtin_bfin_mult_fr1x32", BFIN_BUILTIN_MULT_1X32, MACFLAG_NONE },
5478 { CODE_FOR_flag_mulhi, "__builtin_bfin_mult_fr1x16", BFIN_BUILTIN_MULT_1X16, MACFLAG_T },
5479 { CODE_FOR_flag_mulhi, "__builtin_bfin_multr_fr1x16", BFIN_BUILTIN_MULTR_1X16, MACFLAG_NONE },
5480 { CODE_FOR_flag_mulv2hi, "__builtin_bfin_mult_fr2x16", BFIN_BUILTIN_MULT_2X16, MACFLAG_T },
5481 { CODE_FOR_flag_mulv2hi, "__builtin_bfin_multr_fr2x16", BFIN_BUILTIN_MULTR_2X16, MACFLAG_NONE }
5484 static const struct builtin_description bdesc_1arg[] =
5486 { CODE_FOR_loadbytes, "__builtin_bfin_loadbytes", BFIN_BUILTIN_LOADBYTES, 0 },
5488 { CODE_FOR_ones, "__builtin_bfin_ones", BFIN_BUILTIN_ONES, 0 },
5490 { CODE_FOR_signbitshi2, "__builtin_bfin_norm_fr1x16", BFIN_BUILTIN_NORM_1X16, 0 },
5491 { CODE_FOR_ssneghi2, "__builtin_bfin_negate_fr1x16", BFIN_BUILTIN_NEG_1X16, 0 },
5492 { CODE_FOR_abshi2, "__builtin_bfin_abs_fr1x16", BFIN_BUILTIN_ABS_1X16, 0 },
5494 { CODE_FOR_signbitssi2, "__builtin_bfin_norm_fr1x32", BFIN_BUILTIN_NORM_1X32, 0 },
5495 { CODE_FOR_ssroundsi2, "__builtin_bfin_round_fr1x32", BFIN_BUILTIN_ROUND_1X32, 0 },
5496 { CODE_FOR_ssnegsi2, "__builtin_bfin_negate_fr1x32", BFIN_BUILTIN_NEG_1X32, 0 },
5497 { CODE_FOR_ssabssi2, "__builtin_bfin_abs_fr1x32", BFIN_BUILTIN_ABS_1X32, 0 },
5499 { CODE_FOR_movv2hi_hi_low, "__builtin_bfin_extract_lo", BFIN_BUILTIN_EXTRACTLO, 0 },
5500 { CODE_FOR_movv2hi_hi_high, "__builtin_bfin_extract_hi", BFIN_BUILTIN_EXTRACTHI, 0 },
5501 { CODE_FOR_ssnegv2hi2, "__builtin_bfin_negate_fr2x16", BFIN_BUILTIN_NEG_2X16, 0 },
5502 { CODE_FOR_ssabsv2hi2, "__builtin_bfin_abs_fr2x16", BFIN_BUILTIN_ABS_2X16, 0 }
5505 /* Errors in the source file can cause expand_expr to return const0_rtx
5506 where we expect a vector. To avoid crashing, use one of the vector
5507 clear instructions. */
5509 safe_vector_operand (rtx x, enum machine_mode mode)
5511 if (x != const0_rtx)
5513 x = gen_reg_rtx (SImode);
5515 emit_insn (gen_movsi (x, CONST0_RTX (SImode)));
5516 return gen_lowpart (mode, x);
5519 /* Subroutine of bfin_expand_builtin to take care of binop insns. MACFLAG is -1
5520 if this is a normal binary op, or one of the MACFLAG_xxx constants. */
5523 bfin_expand_binop_builtin (enum insn_code icode, tree exp, rtx target,
5527 tree arg0 = CALL_EXPR_ARG (exp, 0);
5528 tree arg1 = CALL_EXPR_ARG (exp, 1);
5529 rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5530 rtx op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
5531 enum machine_mode op0mode = GET_MODE (op0);
5532 enum machine_mode op1mode = GET_MODE (op1);
5533 enum machine_mode tmode = insn_data[icode].operand[0].mode;
5534 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
5535 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
5537 if (VECTOR_MODE_P (mode0))
5538 op0 = safe_vector_operand (op0, mode0);
5539 if (VECTOR_MODE_P (mode1))
5540 op1 = safe_vector_operand (op1, mode1);
5543 || GET_MODE (target) != tmode
5544 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
5545 target = gen_reg_rtx (tmode);
5547 if ((op0mode == SImode || op0mode == VOIDmode) && mode0 == HImode)
5550 op0 = gen_lowpart (HImode, op0);
5552 if ((op1mode == SImode || op1mode == VOIDmode) && mode1 == HImode)
5555 op1 = gen_lowpart (HImode, op1);
5557 /* In case the insn wants input operands in modes different from
5558 the result, abort. */
5559 gcc_assert ((op0mode == mode0 || op0mode == VOIDmode)
5560 && (op1mode == mode1 || op1mode == VOIDmode));
5562 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
5563 op0 = copy_to_mode_reg (mode0, op0);
5564 if (! (*insn_data[icode].operand[2].predicate) (op1, mode1))
5565 op1 = copy_to_mode_reg (mode1, op1);
5568 pat = GEN_FCN (icode) (target, op0, op1);
5570 pat = GEN_FCN (icode) (target, op0, op1, GEN_INT (macflag));
5578 /* Subroutine of bfin_expand_builtin to take care of unop insns. */
5581 bfin_expand_unop_builtin (enum insn_code icode, tree exp,
5585 tree arg0 = CALL_EXPR_ARG (exp, 0);
5586 rtx op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5587 enum machine_mode op0mode = GET_MODE (op0);
5588 enum machine_mode tmode = insn_data[icode].operand[0].mode;
5589 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
5592 || GET_MODE (target) != tmode
5593 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
5594 target = gen_reg_rtx (tmode);
5596 if (VECTOR_MODE_P (mode0))
5597 op0 = safe_vector_operand (op0, mode0);
5599 if (op0mode == SImode && mode0 == HImode)
5602 op0 = gen_lowpart (HImode, op0);
5604 gcc_assert (op0mode == mode0 || op0mode == VOIDmode);
5606 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
5607 op0 = copy_to_mode_reg (mode0, op0);
5609 pat = GEN_FCN (icode) (target, op0);
5616 /* Expand an expression EXP that calls a built-in function,
5617 with result going to TARGET if that's convenient
5618 (and in mode MODE if that's convenient).
5619 SUBTARGET may be used as the target for computing one of EXP's operands.
5620 IGNORE is nonzero if the value is to be ignored. */
5623 bfin_expand_builtin (tree exp, rtx target ATTRIBUTE_UNUSED,
5624 rtx subtarget ATTRIBUTE_UNUSED,
5625 enum machine_mode mode ATTRIBUTE_UNUSED,
5626 int ignore ATTRIBUTE_UNUSED)
5629 enum insn_code icode;
5630 const struct builtin_description *d;
5631 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
5632 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
5633 tree arg0, arg1, arg2;
5634 rtx op0, op1, op2, accvec, pat, tmp1, tmp2, a0reg, a1reg;
5635 enum machine_mode tmode, mode0;
5639 case BFIN_BUILTIN_CSYNC:
5640 emit_insn (gen_csync ());
5642 case BFIN_BUILTIN_SSYNC:
5643 emit_insn (gen_ssync ());
5646 case BFIN_BUILTIN_DIFFHL_2X16:
5647 case BFIN_BUILTIN_DIFFLH_2X16:
5648 case BFIN_BUILTIN_SUM_2X16:
5649 arg0 = CALL_EXPR_ARG (exp, 0);
5650 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5651 icode = (fcode == BFIN_BUILTIN_DIFFHL_2X16 ? CODE_FOR_subhilov2hi3
5652 : fcode == BFIN_BUILTIN_DIFFLH_2X16 ? CODE_FOR_sublohiv2hi3
5653 : CODE_FOR_ssaddhilov2hi3);
5654 tmode = insn_data[icode].operand[0].mode;
5655 mode0 = insn_data[icode].operand[1].mode;
5658 || GET_MODE (target) != tmode
5659 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
5660 target = gen_reg_rtx (tmode);
5662 if (VECTOR_MODE_P (mode0))
5663 op0 = safe_vector_operand (op0, mode0);
5665 if (! (*insn_data[icode].operand[1].predicate) (op0, mode0))
5666 op0 = copy_to_mode_reg (mode0, op0);
5668 pat = GEN_FCN (icode) (target, op0, op0);
5674 case BFIN_BUILTIN_MULT_1X32X32:
5675 case BFIN_BUILTIN_MULT_1X32X32NS:
5676 arg0 = CALL_EXPR_ARG (exp, 0);
5677 arg1 = CALL_EXPR_ARG (exp, 1);
5678 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5679 op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
5681 || !register_operand (target, SImode))
5682 target = gen_reg_rtx (SImode);
5684 a1reg = gen_rtx_REG (PDImode, REG_A1);
5685 a0reg = gen_rtx_REG (PDImode, REG_A0);
5686 tmp1 = gen_lowpart (V2HImode, op0);
5687 tmp2 = gen_lowpart (V2HImode, op1);
5688 emit_insn (gen_flag_macinit1hi (a1reg,
5689 gen_lowpart (HImode, op0),
5690 gen_lowpart (HImode, op1),
5691 GEN_INT (MACFLAG_FU)));
5692 emit_insn (gen_lshrpdi3 (a1reg, a1reg, GEN_INT (16)));
5694 if (fcode == BFIN_BUILTIN_MULT_1X32X32)
5695 emit_insn (gen_flag_mul_macv2hi_parts_acconly (a0reg, a1reg, tmp1, tmp2,
5696 const1_rtx, const1_rtx,
5697 const1_rtx, const0_rtx, a1reg,
5698 const0_rtx, GEN_INT (MACFLAG_NONE),
5699 GEN_INT (MACFLAG_M)));
5702 /* For saturating multiplication, there's exactly one special case
5703 to be handled: multiplying the smallest negative value with
5704 itself. Due to shift correction in fractional multiplies, this
5705 can overflow. Iff this happens, OP2 will contain 1, which, when
5706 added in 32 bits to the smallest negative, wraps to the largest
5707 positive, which is the result we want. */
5708 op2 = gen_reg_rtx (V2HImode);
5709 emit_insn (gen_packv2hi (op2, tmp1, tmp2, const0_rtx, const0_rtx));
5710 emit_insn (gen_movsibi (gen_rtx_REG (BImode, REG_CC),
5711 gen_lowpart (SImode, op2)));
5712 emit_insn (gen_flag_mul_macv2hi_parts_acconly_andcc0 (a0reg, a1reg, tmp1, tmp2,
5713 const1_rtx, const1_rtx,
5714 const1_rtx, const0_rtx, a1reg,
5715 const0_rtx, GEN_INT (MACFLAG_NONE),
5716 GEN_INT (MACFLAG_M)));
5717 op2 = gen_reg_rtx (SImode);
5718 emit_insn (gen_movbisi (op2, gen_rtx_REG (BImode, REG_CC)));
5720 emit_insn (gen_flag_machi_parts_acconly (a1reg, tmp2, tmp1,
5721 const1_rtx, const0_rtx,
5722 a1reg, const0_rtx, GEN_INT (MACFLAG_M)));
5723 emit_insn (gen_ashrpdi3 (a1reg, a1reg, GEN_INT (15)));
5724 emit_insn (gen_sum_of_accumulators (target, a0reg, a0reg, a1reg));
5725 if (fcode == BFIN_BUILTIN_MULT_1X32X32NS)
5726 emit_insn (gen_addsi3 (target, target, op2));
5729 case BFIN_BUILTIN_CPLX_MUL_16:
5730 case BFIN_BUILTIN_CPLX_MUL_16_S40:
5731 arg0 = CALL_EXPR_ARG (exp, 0);
5732 arg1 = CALL_EXPR_ARG (exp, 1);
5733 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5734 op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
5735 accvec = gen_reg_rtx (V2PDImode);
5738 || GET_MODE (target) != V2HImode
5739 || ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
5740 target = gen_reg_rtx (tmode);
5741 if (! register_operand (op0, GET_MODE (op0)))
5742 op0 = copy_to_mode_reg (GET_MODE (op0), op0);
5743 if (! register_operand (op1, GET_MODE (op1)))
5744 op1 = copy_to_mode_reg (GET_MODE (op1), op1);
5746 if (fcode == BFIN_BUILTIN_CPLX_MUL_16)
5747 emit_insn (gen_flag_macinit1v2hi_parts (accvec, op0, op1, const0_rtx,
5748 const0_rtx, const0_rtx,
5749 const1_rtx, GEN_INT (MACFLAG_W32)));
5751 emit_insn (gen_flag_macinit1v2hi_parts (accvec, op0, op1, const0_rtx,
5752 const0_rtx, const0_rtx,
5753 const1_rtx, GEN_INT (MACFLAG_NONE)));
5754 emit_insn (gen_flag_macv2hi_parts (target, op0, op1, const1_rtx,
5755 const1_rtx, const1_rtx,
5756 const0_rtx, accvec, const1_rtx, const0_rtx,
5757 GEN_INT (MACFLAG_NONE), accvec));
5761 case BFIN_BUILTIN_CPLX_MAC_16:
5762 case BFIN_BUILTIN_CPLX_MSU_16:
5763 case BFIN_BUILTIN_CPLX_MAC_16_S40:
5764 case BFIN_BUILTIN_CPLX_MSU_16_S40:
5765 arg0 = CALL_EXPR_ARG (exp, 0);
5766 arg1 = CALL_EXPR_ARG (exp, 1);
5767 arg2 = CALL_EXPR_ARG (exp, 2);
5768 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5769 op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
5770 op2 = expand_expr (arg2, NULL_RTX, VOIDmode, 0);
5771 accvec = gen_reg_rtx (V2PDImode);
5774 || GET_MODE (target) != V2HImode
5775 || ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
5776 target = gen_reg_rtx (tmode);
5777 if (! register_operand (op1, GET_MODE (op1)))
5778 op1 = copy_to_mode_reg (GET_MODE (op1), op1);
5779 if (! register_operand (op2, GET_MODE (op2)))
5780 op2 = copy_to_mode_reg (GET_MODE (op2), op2);
5782 tmp1 = gen_reg_rtx (SImode);
5783 tmp2 = gen_reg_rtx (SImode);
5784 emit_insn (gen_ashlsi3 (tmp1, gen_lowpart (SImode, op0), GEN_INT (16)));
5785 emit_move_insn (tmp2, gen_lowpart (SImode, op0));
5786 emit_insn (gen_movstricthi_1 (gen_lowpart (HImode, tmp2), const0_rtx));
5787 emit_insn (gen_load_accumulator_pair (accvec, tmp1, tmp2));
5788 if (fcode == BFIN_BUILTIN_CPLX_MAC_16
5789 || fcode == BFIN_BUILTIN_CPLX_MSU_16)
5790 emit_insn (gen_flag_macv2hi_parts_acconly (accvec, op1, op2, const0_rtx,
5791 const0_rtx, const0_rtx,
5792 const1_rtx, accvec, const0_rtx,
5794 GEN_INT (MACFLAG_W32)));
5796 emit_insn (gen_flag_macv2hi_parts_acconly (accvec, op1, op2, const0_rtx,
5797 const0_rtx, const0_rtx,
5798 const1_rtx, accvec, const0_rtx,
5800 GEN_INT (MACFLAG_NONE)));
5801 if (fcode == BFIN_BUILTIN_CPLX_MAC_16
5802 || fcode == BFIN_BUILTIN_CPLX_MAC_16_S40)
5812 emit_insn (gen_flag_macv2hi_parts (target, op1, op2, const1_rtx,
5813 const1_rtx, const1_rtx,
5814 const0_rtx, accvec, tmp1, tmp2,
5815 GEN_INT (MACFLAG_NONE), accvec));
5819 case BFIN_BUILTIN_CPLX_SQU:
5820 arg0 = CALL_EXPR_ARG (exp, 0);
5821 op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
5822 accvec = gen_reg_rtx (V2PDImode);
5823 icode = CODE_FOR_flag_mulv2hi;
5824 tmp1 = gen_reg_rtx (V2HImode);
5825 tmp2 = gen_reg_rtx (V2HImode);
5828 || GET_MODE (target) != V2HImode
5829 || ! (*insn_data[icode].operand[0].predicate) (target, V2HImode))
5830 target = gen_reg_rtx (V2HImode);
5831 if (! register_operand (op0, GET_MODE (op0)))
5832 op0 = copy_to_mode_reg (GET_MODE (op0), op0);
5834 emit_insn (gen_flag_mulv2hi (tmp1, op0, op0, GEN_INT (MACFLAG_NONE)));
5836 emit_insn (gen_flag_mulhi_parts (tmp2, op0, op0, const0_rtx,
5837 const0_rtx, const1_rtx,
5838 GEN_INT (MACFLAG_NONE)));
5840 emit_insn (gen_ssaddhi3_parts (target, tmp2, tmp2, const1_rtx,
5841 const0_rtx, const0_rtx));
5843 emit_insn (gen_sssubhi3_parts (target, tmp1, tmp1, const0_rtx,
5844 const0_rtx, const1_rtx));
5852 for (i = 0, d = bdesc_2arg; i < ARRAY_SIZE (bdesc_2arg); i++, d++)
5853 if (d->code == fcode)
5854 return bfin_expand_binop_builtin (d->icode, exp, target,
5857 for (i = 0, d = bdesc_1arg; i < ARRAY_SIZE (bdesc_1arg); i++, d++)
5858 if (d->code == fcode)
5859 return bfin_expand_unop_builtin (d->icode, exp, target);
5864 #undef TARGET_INIT_BUILTINS
5865 #define TARGET_INIT_BUILTINS bfin_init_builtins
5867 #undef TARGET_EXPAND_BUILTIN
5868 #define TARGET_EXPAND_BUILTIN bfin_expand_builtin
5870 #undef TARGET_ASM_GLOBALIZE_LABEL
5871 #define TARGET_ASM_GLOBALIZE_LABEL bfin_globalize_label
5873 #undef TARGET_ASM_FILE_START
5874 #define TARGET_ASM_FILE_START output_file_start
5876 #undef TARGET_ATTRIBUTE_TABLE
5877 #define TARGET_ATTRIBUTE_TABLE bfin_attribute_table
5879 #undef TARGET_COMP_TYPE_ATTRIBUTES
5880 #define TARGET_COMP_TYPE_ATTRIBUTES bfin_comp_type_attributes
5882 #undef TARGET_RTX_COSTS
5883 #define TARGET_RTX_COSTS bfin_rtx_costs
5885 #undef TARGET_ADDRESS_COST
5886 #define TARGET_ADDRESS_COST bfin_address_cost
5888 #undef TARGET_ASM_INTEGER
5889 #define TARGET_ASM_INTEGER bfin_assemble_integer
5891 #undef TARGET_MACHINE_DEPENDENT_REORG
5892 #define TARGET_MACHINE_DEPENDENT_REORG bfin_reorg
5894 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
5895 #define TARGET_FUNCTION_OK_FOR_SIBCALL bfin_function_ok_for_sibcall
5897 #undef TARGET_ASM_OUTPUT_MI_THUNK
5898 #define TARGET_ASM_OUTPUT_MI_THUNK bfin_output_mi_thunk
5899 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
5900 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
5902 #undef TARGET_SCHED_ADJUST_COST
5903 #define TARGET_SCHED_ADJUST_COST bfin_adjust_cost
5905 #undef TARGET_SCHED_ISSUE_RATE
5906 #define TARGET_SCHED_ISSUE_RATE bfin_issue_rate
5908 #undef TARGET_PROMOTE_PROTOTYPES
5909 #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true
5910 #undef TARGET_PROMOTE_FUNCTION_ARGS
5911 #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_const_tree_true
5912 #undef TARGET_PROMOTE_FUNCTION_RETURN
5913 #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_const_tree_true
5915 #undef TARGET_ARG_PARTIAL_BYTES
5916 #define TARGET_ARG_PARTIAL_BYTES bfin_arg_partial_bytes
5918 #undef TARGET_PASS_BY_REFERENCE
5919 #define TARGET_PASS_BY_REFERENCE bfin_pass_by_reference
5921 #undef TARGET_SETUP_INCOMING_VARARGS
5922 #define TARGET_SETUP_INCOMING_VARARGS setup_incoming_varargs
5924 #undef TARGET_STRUCT_VALUE_RTX
5925 #define TARGET_STRUCT_VALUE_RTX bfin_struct_value_rtx
5927 #undef TARGET_VECTOR_MODE_SUPPORTED_P
5928 #define TARGET_VECTOR_MODE_SUPPORTED_P bfin_vector_mode_supported_p
5930 #undef TARGET_HANDLE_OPTION
5931 #define TARGET_HANDLE_OPTION bfin_handle_option
5933 #undef TARGET_DEFAULT_TARGET_FLAGS
5934 #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT
5936 #undef TARGET_SECONDARY_RELOAD
5937 #define TARGET_SECONDARY_RELOAD bfin_secondary_reload
5939 #undef TARGET_DELEGITIMIZE_ADDRESS
5940 #define TARGET_DELEGITIMIZE_ADDRESS bfin_delegitimize_address
5942 #undef TARGET_CANNOT_FORCE_CONST_MEM
5943 #define TARGET_CANNOT_FORCE_CONST_MEM bfin_cannot_force_const_mem
5945 #undef TARGET_RETURN_IN_MEMORY
5946 #define TARGET_RETURN_IN_MEMORY bfin_return_in_memory
5948 struct gcc_target targetm = TARGET_INITIALIZER;